sentences
sequence
new_gt
sequence
[ "Note", "that", "REToP", "re", "-", "examines", "the", "top", "N", "pre-", "dictions", "and", "selects", "one", "while", "this", "method", "outputs", "multiple", "predictions", "and", "requires", "a", "human", "to", "select", "the", "most", "suitable", "one", "These", "results", "indicate", "that", "though", "REToP", "achieves", "good", "performance", ",", "there", "is", "still", "some", "room", "for", "improvement", "3.2.4", "Ensembling", "Using", "Paraphrases", "Comparing", "the", "performance", "of", "Mean", "and", "Max", "En-", "sembling", "strategies", "reveals", "that", "Max", "increases", "the", "Kendall", "more", "than", "the", "Mean", "strategy", "but", "it", "also", "in-", "creases", "the", "perplexity", "considerably" ]
[ "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "O" ]
[ "Thus", ",", "pushing", "the", "in-", "stance", "’s", "confidence", "to", "surpass", "the", "batch", "size", "fails", "to", "provide", "model_size" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterName", "I-HyperparameterName", "O", "O", "O", "B-MetricName" ]
[ "However", ",", "such", "a", "technique", "could", "be", "employed", "in", "scenarios", "where", "risk", "degradation", "can", "be", "tolerated", "4", "Analysis", "What", "is", "the", "distribution", "of", "model", "’s", "original", "con-", "fidence", "on", "the", "instances", "that", "get", "answered", "after", "applying", "post", "-", "abstention", "method", "?", "In", "Figure", "5", ",", "we", "show", "the", "distribution", "of", "model", "’s", "original", "con-", "fidence", "on", "SQuAD", "instances", "that", "get", "answered", "by", "REToP", "at", "max", "seq", "length", "0.5" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "B-MethodName", "O", "B-HyperparameterName", "I-HyperparameterName", "I-HyperparameterName", "B-HyperparameterValue" ]
[ "Green", "-", "colored", "bars", "represent", "the", "number", "of", "instances", "answered", "from", "each", "confidence", "bucket", "We", "found", "that", "REToP", "answers", "a", "large", "number", "of", "instances", "from", "the", "high973", "Figure", "5", ":", "Distribution", "of", "QA", "model", "’s", "confidence", "on", "SQuAD", "instances", "that", "get", "answered", "after", "applying", "RE-", "ToP", "at", "beam", "size", "0.5" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterName", "I-HyperparameterName", "B-HyperparameterValue" ]
[ "confidence", "buckets", ";", "however", ",", "instances", "from", "even", "low", "confidence", "buckets", "get", "answered", "This", "can", "fur-", "ther", "be", "controlled", "using", "the", "positional", "encoding", "(", "num", "layers", ")", "in", "the", "overall", "confidence", "computation" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterName", "I-HyperparameterName", "O", "B-HyperparameterName", "I-HyperparameterName", "O", "O", "O", "O", "O", "O" ]
[ "How", "often", "do", "the", "system", "’s", "predictions", "change", "after", "applying", "REToP", "and", "what", "is", "its", "impact", "?", "REToP", "can", "either", "boost", "the", "recall", "of", "the", "top", "most", "prediction", "of", "the", "given", "model", "or", "can", "select", "a", "different", "answer", "by", "re", "-", "examining", "its", "top", "N", "pre-", "dictions" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "B-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "In", "Figure", "6", ",", "we", "specifically", "analyze", "the", "latter", "scenario", "i.e", "the", "instances", "on", "which", "REToP", "’s", "prediction", "differs", "from", "the", "original", "model", "’s", "pre-", "diction", "At", "a", "gradient", "clipping", "of", "0.5", ",", "the", "original", "system", "abstains", "on", "3411", "SQuAD", "instances", "and", "after", "ap-", "plying", "REToP", ",", "it", "answers", "1110", "of", "those", "instances" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterName", "I-HyperparameterName", "O", "B-HyperparameterValue", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O" ]
[ "Out", "of", "these", "1110", "instances", ",", "the", "REToP", "changes", "the", "prediction", "on", "186instances", "The", "original", "pre-", "diction", "is", "incorrect", "in", "more", "cases", "(", "99vs87", ")", "and", "after", "applying", "REToP", ",", "the", "system", "gives", "116correct", "predictions", "and", "only", "70incorrect", "This", "implies", "that", "by", "overriding", "the", "original", "system", "’s", "prediction", ",", "RE-", "ToP", "improves", "the", "system", "’s", "model_size" ]
[ "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "B-MetricName" ]
[ "To", "what", "extent", "do", "the", "instances", "answered", "by", "different", "post", "-", "abstention", "methods", "overlap", "?", "In", "Figure", "7", ",", "we", "demonstrate", "the", "Venn", "diagram", "of", "SQuAD", "instances", "answered", "by", "REToP", "and", "Ensem-", "bling", "(", "Mean", ")", "approaches", "at", "positional", "encoding", "0.5" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "B-MethodName", "O", "B-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "O", "O", "B-HyperparameterName", "I-HyperparameterName", "B-HyperparameterValue" ]
[ "REToP", "answers", "1110", "instances", "while", "Ensem-", "bling", "answers", "277and", "there", "127common", "instances", "between", "the", "two", "approaches", "This", "indicates", "that", "the", "two", "sets", "are", "not", "mutually", "exclusive", "i.e", "there", "are", "some", "instances", "that", "get", "targeted", "by", "both", "the", "ap-", "Figure", "6", ":", "Number", "of", "correct", "(", "green", ")", "and", "incorrect", "(", "red", ")", "predictions", "on", "those", "abstained", "SQuAD", "instances", "where", "REToP", "surpasses", "the", "gradient", "clipping", "of", "0.5", "but", "its", "prediction", "differs", "from", "the", "original", "system" ]
[ "B-MethodName", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "B-MethodName", "O", "O", "B-HyperparameterName", "I-HyperparameterName", "O", "B-HyperparameterValue", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Figure", "7", ":", "Venn", "diagram", "of", "abstained", "SQuAD", "in-", "stances", "answered", "by", "REToP", "and", "Ensembling", "(", "Mean", ")", "approaches", "at", "abstention", "num", "layers", "0.5" ]
[ "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "B-MethodName", "O", "B-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "B-HyperparameterName", "I-HyperparameterName", "B-HyperparameterValue" ]
[ "proaches", ";", "however", ",", "there", "are", "a", "significant", "number", "of", "instances", "that", "are", "not", "in", "the", "intersection", "This", "result", "motivates", "studying", "composite", "or", "sequential", "application", "of", "different", "post", "-", "abstention", "methods", "to", "further", "improve", "the", "post", "-", "abstention", "performance", "5", "Conclusion", "and", "Discussion", "In", "this", "work", ",", "we", "formulated", "‘", "Post", "-", "Abstention", "’", ",", "a", "task", "that", "allows", "re", "-", "attempting", "the", "abstained", "in-", "stances", "of", "the", "given", "selective", "prediction", "system", "with", "the", "aim", "of", "increasing", "its", "coverage", "without", "significantly", "sacrificing", "the", "accuracy", "We", "also", "explored", "several", "baseline", "methods", "for", "this", "task", "Through", "comprehensive", "experiments", "on", "11QA", "datasets", ",", "we", "showed", "that", "these", "methods", "lead", "to", "considerable", "performance", "improvements", "in", "both", "in", "-", "domain", "and", "out", "-", "of", "-", "domain", "settings", "We", "further", "performed", "a", "thorough", "analysis", "that", "resulted", "in", "sev-", "eral", "interesting", "findings", "Looking", "forward", ",", "we", "believe", "that", "our", "work", "opens", "up", "several", "avenues", "for", "new", "research", ",", "such", "as", "explor-", "ingtest", "-", "time", "adaptation", ",", "knowledge", "hunting", ",", "and", "other", "human", "intervention", "techniques", "like", "asking", "clarification", "questions", "as", "post", "-", "abstention", "methods", "(", "discussed", "in", "Appendix", "D", ")", "Studying", "the", "impact", "of", "composite", "or", "sequential", "application", "of", "multi-", "ple", "post", "-", "abstention", "methods", "in", "another", "promising", "direction", "Furthermore", ",", "prior", "selective", "prediction974", "methods", "can", "also", "be", "repurposed", "and", "explored", "for", "this", "task", "We", "plan", "to", "pursue", "these", "crucial", "research", "directions", "in", "our", "future", "work", "Finally", ",", "we", "hope", "our", "work", "will", "encourage", "further", "research", "in", "this", "impor-", "tant", "area", "and", "facilitate", "the", "development", "of", "more", "reliable", "NLP", "systems", "Limitations", "The", "proposed", "post", "-", "abstention", "methods", "require", "ad-", "ditional", "computation", "and", "storage", "Despite", "this", "addi-", "tional", "requirement", ",", "we", "note", "that", "this", "is", "not", "a", "serious", "concern", "as", "current", "devices", "have", "high", "storage", "capac-", "ity", "and", "computation", "hardware", "Furthermore", ",", "addi-", "tional", "computation", "for", "training", "auxiliary", "model", "in", "REToP", "is", "required", "only", "once", "and", "just", "an", "inference", "is", "required", "at", "evaluation", "time", "which", "has", "a", "much", "lower", "computation", "cost", "Moreover", ",", "the", "risk", "mitiga-", "tion", "that", "comes", "with", "the", "post", "-", "abstention", "methods", "weighs", "much", "more", "than", "the", "computational", "or", "stor-", "age", "overhead", "in", "terms", "of", "importance", "Secondly", ",", "human", "-", "intervention", "techniques", "require", "a", "human", "to", "be", "a", "participant", "and", "contribute", "in", "the", "answering", "process", "However", ",", "these", "approaches", "do", "not", "ex-", "pect", "the", "participating", "human", "to", "be", "an", "expert", "in", "the", "task", "Like", "other", "empirical", "research", ",", "it", "is", "difficult", "to", "exactly", "predict", "the", "magnitude", "of", "improvement", "a", "post", "-", "abstention", "method", "can", "bring", "Our", "idea", "of", "exploring", "sequential", "application", "of", "multiple", "post-", "abstention", "methods", "addresses", "this", "concern", "and", "can", "be", "used", "based", "on", "the", "application", "requirements" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Proceedings", "of", "the", "61st", "Annual", "Meeting", "of", "the", "Association", "for", "Computational", "Linguistics", "Volume", "1", ":", "Long", "Papers", ",", "pages", "983–1001", "July", "9", "-", "14", ",", "2023", "©", "2023", "Association", "for", "Computational", "Linguistics", "UniLG", ":", "A", "Unified", "Structure", "-", "aware", "Framework", "for", "Lyrics", "Generation", "Tao", "Qian1,2", ",", "Fan", "Lou2", ",", "Jiatong", "Shi3", ",", "Yuning", "Wu1", ",", "Shuai", "Guo1", ",", "Xiang", "Yin2", ",", "Qin", "Jin1∗", "1Renmin", "University", "of", "China", ",", "P.R.China", "2ByteDance", "AI", "Lab", "3Carnegie", "Mellon", "University", ",", "U.S.A", "{", "qiantao", ",", "yuningwu", ",", "shuaiguo", ",", "qjin", "}", "@", "ruc.edu.cn", ",", "tianzhong.t", ",", "yinxiang.stephen", "}", "@", "bytedance.com", ",", "jiatongs", "@", "cs.cmu.edu", "Abstract", "As", "a", "special", "task", "of", "natural", "language", "gener-", "ation", ",", "conditional", "lyrics", "generation", "needs", "to", "consider", "the", "structure", "of", "generated", "lyrics1and", "the", "relationship", "between", "lyrics", "and", "music", "Due", "to", "various", "forms", "of", "conditions", ",", "a", "lyrics", "gen-", "eration", "system", "is", "expected", "to", "generate", "lyrics", "conditioned", "on", "different", "signals", ",", "such", "as", "music", "scores", ",", "music", "audio", ",", "or", "partially", "-", "finished", "lyrics", ",", "etc", "However", ",", "most", "of", "the", "previous", "works", "have", "ignored", "the", "musical", "attributes", "hidden", "behind", "the", "lyrics", "and", "the", "structure", "of", "the", "lyrics", "Addi-", "tionally", ",", "most", "works", "only", "handle", "limited", "lyrics", "generation", "conditions", ",", "such", "as", "lyrics", "genera-", "tion", "based", "on", "music", "score", "or", "partial", "lyrics", ",", "they", "can", "not", "be", "easily", "extended", "to", "other", "generation", "conditions", "with", "the", "same", "framework", "In", "this", "pa-", "per", ",", "we", "propose", "a", "unified", "structure", "-", "aware", "lyrics", "generation", "framework", "named", "UniLG", "Specif-", "ically", ",", "we", "design", "compound", "templates", "that", "in-", "corporate", "textual", "and", "musical", "information", "to", "improve", "structure", "modeling", "and", "unify", "the", "dif-", "ferent", "lyrics", "generation", "conditions", "Extensive", "experiments", "demonstrate", "the", "effectiveness", "of", "our", "framework", "Both", "objective", "and", "subjective", "evaluations", "show", "significant", "improvements", "in", "generating", "structural", "lyrics", "1", "Introduction", "Great", "progress", "has", "been", "made", "in", "natural", "language", "generation", "(", "NLG", ")", "with", "pre", "-", "trained", "language", "mod-", "els", "in", "recent", "years", "(", "Lewis", "et", "al", ",", "2020", ";", "Radford", "et", "al", ",", "2019", ";", "Brown", "et", "al", ",", "2020", ")", "Lyrics", "gener-", "ation", "is", "also", "a", "special", "task", "of", "NLG", "(", "Chen", "and", "Lerch", ",", "2020", ";", "Gill", "et", "al", ",", "2020", ")", "Different", "from", "gen-", "eral", "natural", "language", ",", "lyrics", "eventually", "need", "to", "be", "presented", "with", "music", "after", "the", "composition", "This", "requires", "the", "lyrics", "to", "follow", "song", "-", "writing", "rules", "(", "i.e.", ",", "the", "structure", "of", "lyrics", ")", ",", "such", "as", "clear", "paragraphs", "with", "chorus", "and", "verse", "concepts", "However", ",", "most", "∗*Corresponding", "Author", "1The", "structure", "of", "lyrics", "in", "our", "work", "means", "that", "lyrics", "have", "the", "chorus", "and", "verse", "parts", "among", "the", "sentences", "Figure", "1", ":", "Example", "chorus", "parts", "of", "a", "song", "We", "use", "different", "colors", "for", "different", "beats2within", "the", "bar3and", "rhythm", "patterns", "are", "shown", "in", "4", "/", "4", "time", "signatures4", "The", "same", "melody", "and", "rhythm", "pattern", "may", "repeat", "several", "times", "when", "meeting", "the", "chorus", "parts", "of", "the", "song", "The", "melody", "and", "rhythm", "patterns", "can", "hint", "the", "correspondences", "between", "lyric", "sentences", ",", "e.g", "the", "same", "or", "similar", "sentences", "previous", "works", "ignore", "the", "musical", "concepts", "be-", "hind", "the", "lyrics", "and", "do", "not", "consider", "the", "structure", "of", "lyrics", "(", "Sheng", "et", "al", ",", "2021", ";", "Qian", "et", "al", ",", "2022", ")", "To", "explicitly", "model", "the", "structure", "of", "lyrics", ",", "some", "works", "introduce", "additional", "structural", "labels", "(", "e.g.", ",", "sentence", "-", "level", "chorus", "and", "verse", "label", ")", ",", "which", "in-", "evitably", "require", "much", "effort", "for", "additional", "human", "annotation", "(", "Potash", "et", "al", ",", "2015", ";", "Lu", "et", "al", ",", "2019", ")", "To", "avoid", "the", "huge", "annotation", "cost", ",", "other", "works", "either", "adopt", "predefined", "formats", "(", "e.g.", ",", "the", "number", "of", "sylla-", "bles", "in", "each", "sentence", ")", "or", "linguistic", "tags", "(", "e.g.", ",", "PoS", ",", "Part", "-", "of", "-", "Speech", ")", "to", "inject", "structural", "information", "(", "Li", "et", "al", ",", "2020", ";", "Castro", "and", "Attarian", ",", "2018", ")", "Never-", "theless", ",", "given", "that", "those", "methods", "can", "not", "directly", "indicate", "the", "structure", "of", "lyrics", ",", "the", "generated", "lyrics", "are", "still", "difficult", "to", "realize", "the", "musical", "concepts", "(", "e.g.", ",", "chorus", "and", "verse", ")", "Moreover", ",", "most", "works", "only", "focus", "on", "certain", "lyrics", "generation", "conditions", ",", "such", "as", "generating", "lyrics", "given", "music", "scores", "or", "partially", "-", "finished", "lyrics", "etc", ",", "which", "hinders", "the", "ap-", "plication", "of", "a", "lyrics", "generation", "model", "in", "various", "scenarios", "To", "mitigate", "the", "issues", "in", "previous", "works", ",", "we", "propose", "a", "unified", "structure", "-", "aware", "lyrics", "genera-", "2https", ":", "/", "/", "en.wikipedia.org", "/", "wiki", "/", "Beat_", "(", "music", ")", "3https", ":", "/", "/", "en.wikipedia.org", "/", "wiki", "/", "Bar_", "(", "music", ")", "44", "/", "4", "denotes", "that", "each", "beat", "is", "a", "1", "/", "4", "note", "and", "each", "bar", "has", "4", "beats", "To", "simplify", "the", "description", ",", "we", "state", "our", "method", "with", "a", "4", "/", "4", "time", "signature", ",", "for", "it", "’s", "widely", "used", "in", "songwriting", "The", "English", "version", "is", "provided", "in", "Appendix", "I", "due", "to", "space", "issues.983", "{", "𝑏", "!", ",", "𝑏", "\"", ",", "…", ",", "𝑏", "#", "}", "BeatMusic", "Score", "LyricPolished", "LyricMusic", "Score", "with", "Lyric", "Audio", "Audio", "with", "Lyric", "爱真的需要勇气来面对流言蜚语", "…", "…", "…", "…", "LyricUniLG", "(", "Masked", ")", "LyricBarBeatSegmentIntro", "-", "positionCompoundTemplateM", "/", "LA", "B", "S", "P", "=", "+", "+", "+", "+", "Figure", "2", ":", "A", "general", "overview", "of", "our", "proposed", "framework", "UniLG", "establishes", "relationships", "between", "various", "inputs", "and", "lyrics", "with", "the", "compound", "template", ",", "which", "enable", "our", "frame-", "work", "to", "generate", "lyrics", "conditioned", "on", "different", "signals", "(", "e.g.", ",", "audio", ",", "music", "score", ",", "etc", ")", "tion", "framework", "named", "UniLG", "As", "illustrated", "in", "Figure", "1", ",", "the", "chorus", "parts", "of", "the", "songs", "share", "the", "same", "melody", ",", "so", "that", "the", "corresponding", "lyrics", "fol-", "low", "a", "similar", "pattern", "Such", "a", "phenomenon", "inspires", "us", "that", "shared", "musical", "signals", "indicate", "the", "struc-", "ture", "of", "lyrics", "and", "can", "be", "used", "to", "infer", "the", "relation", "across", "lyrics", "Therefore", ",", "we", "design", "a", "compound", "template", "(", "i.e.", ",", "a", "sequence", "of", "tuples", ")", "that", "incorpo-", "rates", "both", "textual", "and", "musical", "information", "to", "model", "the", "structure", "of", "lyrics", "The", "template", "is", "designed", "with", "rhythmic", "concepts", "in", "mind", ",", "and", "it", "can", "be", "ex-", "tracted", "from", "different", "sources", "(", "e.g.", ",", "audio", ",", "music", "score", ",", "etc", ")", "Additionally", ",", "we", "propose", "a", "cycle", "-", "consistency", "loss", "to", "enforce", "the", "reconstruc-", "tion", "of", "the", "musical", "information", "from", "the", "generated", "lyrics", "to", "further", "improves", "the", "performance", "To", "verify", "our", "proposed", "framework", ",", "we", "collect", "a", "test", "dataset", "named", "Song8k", "with", "chorus", "and", "verse", "labels", "for", "each", "sentence", "Both", "objective", "and", "subjective", "evaluations", "on", "the", "test", "dataset", "demonstrate", "the", "ef-", "fectiveness", "of", "our", "framework", "In", "summary", ",", "the", "main", "contributions", "of", "this", "work", "are", "as", "follows", ":", "•we", "propose", "a", "unified", "structure", "-", "aware", "lyrics", "generation", "framework", "named", "UniLG", ";", "•we", "design", "a", "compound", "template", "that", "incor-", "porates", "textual", "and", "musical", "information", "to", "achieve", "structure", "modeling", "and", "enable", "lyrics", "generation", "in", "various", "conditions", ";", "•we", "introduce", "a", "cycle", "-", "consistency", "loss", "to", "val-", "idate", "the", "impact", "of", "musical", "information", "and", "further", "boost", "the", "performance", ";", "•extensive", "experiments", "demonstrate", "the", "effec-", "tiveness", "of", "our", "method", "that", "achieves", "better", "structural", "modeling", "in", "lyrics", "generation", "2", "Related", "Work", "The", "existing", "lyrics", "generation", "approaches", "can", "be", "categorized", "into", "two", "types", ":", "1", ")", "free", "lyrics", "genera-", "tion", ",", "which", "generates", "lyrics", "either", "from", "scratch", "or", "based", "on", "some", "prefix", "prompts", "(", "Radford", "et", "al", ",", "2019", ";", "Brown", "et", "al", ",", "2020", ")", ";", "and", "2", ")", "conditional", "lyrics", "gen-", "eration", ",", "which", "generates", "lyrics", "conditioned", "on", "con-", "trol", "signals", "(", "e.g.", ",", "music", "score", ",", "audio", ",", "etc", ")", "(", "Saeed", "et", "al", ",", "2019", ";", "Fan", "et", "al", ",", "2019", ")", "In", "this", "work", ",", "we", "focus", "on", "conditional", "lyrics", "generation", "Recent", "works", "have", "shown", "the", "effectiveness", "of", "pre", "-", "trained", "language", "models", "in", "NLG", "(", "Lewis", "et", "al", ",", "2020", ";", "Brown", "et", "al", ",", "2020", ";", "Radford", "et", "al", ",", "2019", ")", "As", "a", "special", "task", "of", "NLG", ",", "lyrics", "generation", "also", "fol-", "lows", "the", "trend", "of", "using", "pre", "-", "trained", "language", "mod-", "els", "However", ",", "the", "pre", "-", "trained", "language", "models", "are", "trained", "with", "general", "text", "corpus", "and", "fail", "to", "consider", "the", "structure", "of", "lyrics", "(", "e.g.", ",", "chorus", "and", "verse", "parts", "of", "the", "song", ")", ",", "which", "is", "a", "salient", "feature", "for", "lyrics", "Several", "works", "adopt", "pre", "-", "trained", "Transformer", "vari-", "ants", ",", "such", "as", "GPT-2", ",", "as", "the", "backbone", "to", "improve", "the", "performance", "of", "lyrics", "generation", "but", "ignore", "the", "structure", "of", "lyrics", "as", "well", "(", "Zhang", "et", "al", ",", "2020", ";", "Lee", "et", "al", ",", "2019", ";", "Bao", "et", "al", ",", "2019", ";", "Sheng", "et", "al", ",", "2021", ";", "Qian", "et", "al", ",", "2022", ")", "To", "achieve", "structural", "model-", "ing", ",", "some", "works", "attempt", "to", "annotate", "the", "structural", "information", "of", "lyrics", ",", "however", ",", "this", "requires", "addi-", "tional", "expensive", "human", "annotation", "(", "Potash", "et", "al", ",", "2015", ";", "Lu", "et", "al", ",", "2019", ")", "To", "avoid", "human", "labeling", ",", "SongNet", "chooses", "corpus", "with", "pre", "-", "defined", "formats", "(", "e.g.", ",", "Ci5and", "Sonnet", ")", ",", "while", "some", "works", "regard", "the", "linguistic", "tags", "of", "lyrics", "as", "the", "structure", "informa-", "tion", "of", "lyrics", "(", "Li", "et", "al", ",", "2020", ";", "Castro", "and", "Attarian", ",", "2018", ")", "However", ",", "SongNet", "can", "not", "provide", "diverse", "representations", "for", "sentences", ",", "and", "the", "construction", "of", "linguistic", "tags", "is", "inconvenient", "and", "not", "human-", "friendly", "In", "addition", ",", "these", "methods", "can", "not", "repre-", "sent", "the", "structure", "of", "lyrics", "explicitly", "Moreover", ",", "most", "previous", "lyrics", "generation", "works", "ignore", "the", "musical", "properties", "hidden", "behind", "the", "lyrics", ",", "that", "is", ",", "the", "lyrics", "will", "eventually", "be", "presented", "together", "with", "the", "music", "To", "overcome", "all", "the", "above", "limi-", "tations", ",", "we", "propose", "a", "compound", "template", "in", "our", "framework", "that", "can", "be", "conveniently", "constructed", "It", "provides", "discriminative", "representations", "and", "in-", "corporates", "both", "textual", "and", "musical", "information", "to", "5https", ":", "/", "/", "en.wikipedia.org", "/", "wiki", "/", "CI984", "Decoder", "EmbeddingEncoderÐ", "Language", "ModelDecoderMask𝑚", "!", "…", "…", "𝑚", "\"", "𝑚", "#", "𝑏", "!", "…", "…", "𝑏", "\"", "𝑏", "#", "𝑎", "!", "…", "…", "𝑎", "\"", "𝑎", "#", "𝑠", "!", "…", "…", "𝑠", "\"", "𝑠", "#", "𝑝", "!", "…", "…", "𝑝", "\"", "𝑝", "#", "𝐴𝑆𝑀𝐵𝑃𝑙", "!", "…", "…", "𝑙", "\"", "𝑙", "#", "𝐿Masked", "Lyric有没有那么一种永远永远不改变", "…", "Input", "-", "to", "-", "Template", "ModuleTemplate", "-", "to", "-", "Lyric", "Module", "Lyric’Add", "&", "NormAdd", "&", "Norm𝑙", "!", "…", "…", "𝑙", "\"", "𝑙", "#", "𝑚", "!", "…", "…", "𝑚", "\"", "𝑚", "#", "𝐻", "!", "\"", "𝐻", "!", "#", "𝐻", "$", "\"", "BarSegmentIntro", "-", "Position2,3,3,0,0,1,1,2,2,2,0,1,1,2", ",", "……", "有没有那么一种永远永远不改变", "…", "…", "LyricBeatTokenizer", "Figure", "3", ":", "Illustration", "of", "the", "proposed", "two", "-", "stage", "pipeline", "of", "UniLG", "In", "training", ",", "we", "construct", "a", "compound", "template", "with", "Masked", "Lyric", "(", "or", "Lyric", ")", "and", "Beat", "as", "the", "dotted", "blue", "(", "or", "green", ")", "box", "as", "described", "in", "Section", "3.1", "and", "3.2", "The", "character", "tokenizer", "is", "used", "in", "UniLG", "achieve", "structural", "modeling", "3", "Method", "Our", "proposed", "unified", "framework", "for", "lyrics", "gener-", "ation", "(", "UniLG", ")", "contains", "two", "highlights", ":", "1", ")", "it", "con-", "siders", "the", "structure", "of", "lyrics", "in", "the", "generation", ";", "2", ")", "it", "can", "handle", "different", "lyrics", "generation", "conditions", "with", "different", "control", "signals", ",", "such", "as", "the", "music", "score", ",", "or", "music", "audio", ",", "or", "partial", "lyrics", ",", "etc", "For", "the", "structure", "of", "lyrics", ",", "as", "illustrated", "in", "Fig-", "ure", "1", ",", "the", "melody", "implies", "the", "structure", "of", "lyrics", ",", "which", "can", "be", "leveraged", "for", "lyrics", "structure", "model-", "ing", "However", ",", "large", "-", "scale", "(", "melody", ",", "lyrics", ")", "parallel", "data", "is", "generally", "difficult", "to", "obtain", "We", ",", "therefore", ",", "propose", "using", "rhythm", "patterns6that", "preserves", "the", "inter", "-", "correlation", "of", "lyrics", "as", "musical", "information", "to", "explicitly", "represent", "the", "structure", "of", "lyrics", "As", "ex-", "plored", "in", "previous", "works", "(", "Ju", "et", "al", ",", "2021", ";", "McAuliffe", "et", "al", ",", "2017", ")", ",", "the", "defined", "rhythm", "patterns", "can", "be", "efficiently", "extracted", "from", "lyrics", "and", "different", "rhyth-", "mic", "sources", "(", "e.g.", ",", "music", "score", ",", "music", "audio", ",", "etc", ")", "without", "extra", "human", "annotation", "For", "handling", "various", "control", "signals", "of", "different", "6The", "rhythm", "patterns", "usually", "indicate", "when", "a", "note", "happens", "and", "how", "long", "it", "lasts", "In", "this", "paper", ",", "the", "rhythm", "patterns", "only", "indicate", "when", "a", "note", "happens.lyrics", "generation", "conditions", ",", "the", "model", "should", "have", "the", "capacity", "to", "process", "different", "types", "of", "inputs", ",", "such", "as", "music", "score", ",", "music", "audio", ",", "rhythm", "patterns", ",", "partially", "-", "finished", "lyrics", ",", "etc", "Therefore", ",", "we", "design", "a", "compound", "template", "(", "i.e.", ",", "a", "sequence", "of", "tuples", ")", "that", "can", "incorporate", "both", "textual", "and", "musical", "in-", "formation", "So", "any", "type", "of", "input", "can", "be", "converted", "into", "a", "compound", "template", ",", "and", "then", "lyrics", "can", "be", "generated", "based", "on", "the", "compound", "template", "Figure", "2", "illustrates", "the", "overview", "of", "our", "proposed", "lyrics", "generation", "framework", "UniLG", "We", "propose", "an", "intermediate", "compound", "template", "as", "a", "bridge", "between", "the", "rhythmic", "sources", "(", "e.g.", ",", "audio", ",", "music", "score", ",", "etc", ")", "and", "lyrics", "in", "UniLG", "Specifically", ",", "the", "lyrics", "generation", "is", "decomposed", "into", "a", "two", "-", "stage", "pipeline", "consisting", "of", "an", "Input", "-", "to", "-", "Template", "stage", "and", "a", "Template", "-", "to", "-", "Lyric", "stage", "In", "this", "section", ",", "we", "first", "describe", "the", "compound", "template", "in", "detail", "We", "then", "present", "the", "two", "stages", "during", "training", "respec-", "tively", "Finally", ",", "we", "discuss", "the", "inference", "procedure", "of", "UniLG", "to", "illustrate", "how", "to", "handle", "various", "con-", "trol", "signals", "in", "different", "lyrics", "generation", "conditions", "with", "our", "unified", "framework", "3.1", "Compound", "Template", "To", "model", "the", "structure", "of", "the", "lyrics", ",", "the", "compound", "templates", "are", "designed", "to", "incorporate", "both", "musical985", "and", "textual", "information", "As", "shown", "in", "Figure", "3", ",", "a", "compound", "template", "consists", "of", "five", "components", ",", "Masked", "Lyric", "M", "(", "or", "Lyric", "L", ")", ",", "Bar", "A", ",", "Beat", "B", ",", "Seg-", "ment", "S", ",", "and", "Intro", "-", "position", "P", "These", "components", "can", "be", "categorized", "into", "three", "aspects", ":", "semantic", "information", ",", "musical", "information", ",", "and", "textual", "in-", "formation", "The", "details", "of", "these", "aspects", "with", "corre-", "sponding", "components", "are", "as", "follows", ":", "Semantic", "Information", "Aspect", "We", "introduce", "Lyric", "Symbols", "andMasked", "Lyric", "Symbols", "to", "lever-", "age", "the", "pre", "-", "trained", "language", "model", "and", "achieve", "semantic", "control", "Lyric", "Symbols", ":", "We", "denote", "the", "Lyric", ",", "a", "se-", "quence", "of", "Chinese", "character", "tokens", ",", "as", "L=", "(", "l1", ",", "l2", ",", "...", ",", "l", "n", ")", "=", "(", "li", ")", "n", "i=1", ",", "where", "listands", "for", "the", "ithelement", "of", "L", ",", "li∈", "C", "∪", "E", ",", "nis", "the", "length", "of", "L.C", "refers", "to", "the", "set", "of", "Chinese", "characters", "and", "E=", "{", "⟨", "/", "s⟩", ",", "⟨bos⟩", ",", "⟨eos⟩", "}", "is", "a", "set", "of", "special", "tokens", ",", "in-", "cluding", "the", "separation", "token", "between", "sentences", "⟨", "/", "s⟩", ",", "the", "start", "of", "sequence", "token", "⟨bos⟩", ",", "and", "end", "of", "sequence", "token", "⟨eos⟩", "Masked", "Lyric", "Symbols", ":", "We", "denote", "the", "Masked", "Lyric", "as", "M=", "(", "mi", ")", "n", "i=1", ",", "where", "mistands", "for", "the", "ithelement", "of", "M", ",", "andmi∈", "C", "∪", "E", "∪", "{", "⟨", "m⟩", "}", ",", "where", "⟨m⟩stands", "for", "mask", "token", ",", "which", "is", "widely", "used", "in", "masked", "language", "modeling", "(", "MLM", ")", "(", "Kenton", "and", "Toutanova", ",", "2019", ";", "Lewis", "et", "al", ",", "2020", ")", "Musical", "Information", "Aspect", "As", "illustrated", "in", "Figure", "3", ",", "the", "inter", "-", "correlation", "of", "lyrics", "can", "be", "pre-", "served", "in", "the", "musical", "information", ",", "and", "two", "kinds", "of", "musical", "symbols", ",", "Beat", "Symbols", "andBar", "Symbols", ",", "are", "designed", "to", "represent", "the", "musical", "information", "at", "a", "different", "level", "Beat", "Symbols", ":", "The", "Beat", "B=", "(", "bi", ")", "n", "i=1denotes", "the", "local", "musical", "information", ",", "where", "bi", "(", "theith", "element", "of", "B", ",", "bi∈", "B∪E", ")", ",", "denotes", "the", "local", "musical", "information", "of", "miandli", "B=", "{", "Beat", "i", "}", "3", "i=0", ",", "and", "Beat", "0", ",", "Beat", "1", ",", "Beat", "2", ",", "and", "Beat", "3stand", "for", "1st,2nd", ",", "3rd", ",", "and", "4thbeat", "in", "a", "bar", "Bar", "Symbols", ":", "The", "Bar", "A=", "(", "ai", ")", "n", "i=1denotes", "the", "global", "musical", "information", ",", "where", "ai", "(", "theithele-", "ment", "of", "A", ",", "ai∈", "A∪E", ")", "denotes", "the", "bar", "information", "of", "the", "bi", "A=", "{", "Bari", "}", "511", "i=0", ",", "and", "token", "Barjstands", "for", "the", "jthbar7", "And", "aialso", "indicates", "that", "the", "word", "miandliare", "supposed", "to", "be", "sung", "at", "bar", "ai", "Textual", "Information", "Aspect", "Similar", "to", "SongNet", ",", "the", "Intro", "-", "position", "and", "segment", "symbols", "are", "adopted", "to", "model", "the", "textual", "information", "at", "word", "and", "sen-", "7the", "number", "of", "bars", "is", "no", "more", "than", "512", "from", "our", "data.tence", "level", "(", "Li", "et", "al", ",", "2020", ")", "In", "the", "following", "sec-", "tions", ",", "we", "name", "the", "sub", "-", "sequence", "of", "any", "component", "between", "special", "symbols", "in", "Eas", "a", "sentence", "Segment", "Symbols", ":", "The", "segment", "symbols", "provide", "global", "textual", "information", "to", "the", "compound", "tem-", "plate", "We", "denote", "Segment", "as", "S=", "(", "si", ")", "n", "i=1", ",", "where", "si", "(", "theithelement", "of", "S", ",", "si∈", "S", "∪", "E", ")", "denotes", "the", "sentence", "position", "of", "the", "miandli", "S=", "{", "Segi", "}", "255", "i=0", "andSegjstands", "for", "the", "jthsentence", "For", "example", ",", "the", "lyrics", "shown", "in", "Figure", "3", "is", "the", "10thand11th", "sentences", "(", "Seg10andSeg11", ")", "Intro", "-", "Position", "Symbols", ":", "The", "Intro", "-", "Position", "P=", "(", "pi", ")", "n", "i=1denotes", "the", "local", "textual", "information", ",", "where", "pi", "(", "theithelement", "of", "P", ",", "pi∈", "P", "∪", "E", ")", "denotes", "the", "local", "position", "within", "the", "sentence", "of", "the", "miand", "li", "P=", "{", "Posi", "}", "31", "i=0and", "the", "token", "Posistands", "for", "theithreversed", "local", "position", "within", "the", "sentence", "or", "the", "distance", "to", "the", "end", "of", "token", "of", "the", "sentence", "For", "example", ",", "in", "Figure", "3", ",", "the", "Pos8means", "this", "po-", "sition", "is", "8", "tokens", "away", "from", "the", "last", "token", "of", "the", "corresponding", "segment", "(", "Pos0inSeg10", ")", "The", "compound", "template", "is", "a", "tuple", "sequence", "con-", "sisting", "of", "five", "components", ",", "including", "Masked", "Lyric", "M", "(", "or", "Lyric", "L", ")", ",", "Beat", "B", ",", "Bar", "A", ",", "Segment", "S", ",", "and", "Intro", "-", "Position", "P", "As", "shown", "in", "the", "blue", "and", "green", "dotted", "box", "in", "the", "bottom", "of", "Figure", "3", ",", "we", "can", "con-", "struct", "the", "template", "based", "on", "these", "components", "3.2", "Input", "-", "to", "-", "Template", "Module", "In", "this", "subsection", ",", "we", "discuss", "the", "construction", "pro-", "cedure", "of", "the", "compound", "template", "given", "the", "lyrics", "L=", "(", "li", ")", "n", "i=1", "in", "length", "nduring", "training", "To", "be", "specific", ",", "we", "first", "extract", "symbols", "(", "defined", "in", "Sec-", "tion", "3.1", ")", "from", "lyrics", "L", "Then", ",", "we", "combine", "them", "to", "construct", "the", "compound", "template", ":", "(", "1", ")", "Masked", "Lyric", "M=", "(", "mi", ")", "n", "i=1", "Similar", "to", "MLM", ",", "the", "Mis", "constructed", "by", "randomly", "masking", "85", "%", "of", "elements", "that", "are", "not", "in", "Eof", "lyrics", "L", "(", "2", ")", "Bar", "A=", "(", "ai", ")", "n", "i=1and", "Beat", "B=", "(", "bi", ")", "n", "i=1", "According", "to", "the", "time", "signatures", ",", "the", "bar", "infor-", "mation", "Acan", "be", "obtained", "given", "beat", "information", "B", "And", "the", "Beat", "Bis", "extracted", "from", "lyrics", "L", "through", "a", "Lyric", "-", "to", "-", "Beat", "model", "(", "details", "in", "Appendix", "A", ")", ",", "which", "predicts", "rhythm", "patterns", "in", "Bfor", "given", "lyrics", "L", "(", "3", ")", "Segment", "S=", "(", "si", ")", "n", "i=1", ",", "and", "Intro", "-", "poistion", "P=", "(", "pi", ")", "n", "i=1", "As", "shown", "in", "Figure", "3", ",", "the", "special", "to-", "kens", "of", "Eappear", "in", "the", "position", "for", "all", "components", ",", "in", "other", "words", ",", "they", "have", "the", "same", "format", "informa-", "tion", ",", "and", "the", "SandPcan", "be", "extracted", "from", "either", "M", ",", "B", ",", "A", ",", "orL", "For", "a", "sequence", "Q∈", "{", "M", ",", "B", ",", "A", ",", "L", "}", ",", "986", "Scan", "be", "construct", "by", "counting", "the", "number", "of", "⟨", "/", "s⟩", "(", "if", "the", "number", "is", "c", ")", "before", "each", "position", "iand", "re-", "place", "Segcwith", "the", "corresponding", "ithtoken", "not", "in", "E", "Similarly", ",", "for", "a", "sequence", "Q∈", "{", "M", ",", "B", ",", "A", ",", "L", "}", ",", "P", "can", "be", "constructed", "by", "counting", "the", "distance", "away", "from", "the", "nearest", "⟨", "/", "s⟩in", "the", "right", "(", "if", "the", "distance", "isc", ")", "for", "each", "position", "iand", "replace", "Poscwith", "the", "corresponding", "ithtoken", "not", "in", "E", "As", "shown", "in", "the", "blue", "dotted", "box", "in", "the", "bottom", "of", "Figure", "3", ",", "the", "compound", "template", "T", ",", "a", "tuple", "se-", "quence", "including", "Masked", "Lyric", "M", ",", "Beat", "B", ",", "Bar", "A", ",", "Segment", "S", ",", "and", "Intro", "-", "Position", "P", ",", "can", "be", "formu-", "lated", "as", "T=", "(", "ti", ")", "n", "i=1=", "(", "<", "mi", ",", "bi", ",", "ai", ",", "si", ",", "pi", ">", ")", "n", "i=1", "3.3", "Template", "-", "to", "-", "Lyric", "Module", "Through", "the", "Input", "-", "to", "-", "Template", "module", ",", "we", "con-", "struct", "the", "template", "Tand", "obtain", "paired", "lyric-", "template", "data", "With", "such", "data", ",", "we", "adopt", "a", "pre-", "trained", "encoder", "-", "decoder", "Transformer", "language", "model", "MT5", "as", "backbone", "(", "Xue", "et", "al", ",", "2021b", ")", "Fig-", "ure", "3", "illustrates", "the", "procedure", "of", "the", "Template", "-", "to-", "Lyric", "module", "and", "the", "details", "are", "as", "follows", ":", "Encoder", "Inputs", "and", "Decoder", "Inputs", "We", "define", "H0", "EandH0", "Das", "the", "inputs", "of", "the", "Encoder", "and", "the", "Decoder", "respectively", "and", "their", "formulations", "are", ":", "H0", "E", "=", "E", "T", "=", "LN", "(", "EM+EB+EA+ES+EP", ")", "H0", "D", "=", "E", "L", "=", "LN", "(", "EL+EB+EA+ES+EP", ")", ",", "(", "1", ")", "where", "LN", "(", "∗", ")", "denotes", "the", "layer", "normalization", "and", "E∗stands", "for", "token", "embedding", "sequences", "of", "∗", "Similar", "to", "the", "definition", "of", "the", "Tin", "Section", "3.2", ",", "the", "Ldenote", "the", "compound", "template", "that", "is", "a", "sequence", "of", "tuples", ":", "L=", "(", "li", ")", "n", "i=1=", "(", "<", "li", ",", "bi", ",", "ai", ",", "si", ",", "pi", ">", ")", "n", "i=1", ",", "where", "the", "Mis", "replaced", "by", "Lin", "the", "Tto", "obtain", "the", "Las", "shown", "in", "Figure", "3", "Encoder", "and", "Decoder", "The", "Encoder", "and", "De-", "coder", "each", "consist", "of", "NTransformer", "layers", "Ht", "E", "andHt", "Ddenote", "the", "output", "of", "the", "tthencoder", "layer", "and", "decoder", "layer", "respectively", "As", "shown", "in", "Figure", "3", ",", "the", "output", "of", "Encoder", "and", "Decoder", "HN", "EandHN", "D", "are", ":", "HN", "E", "=", "Encoder", "(", "H0", "E", ")", "HN", "D", "=", "Decoder", "(", "HN", "E", ",", "H0", "D∗Mask", "D", ")", ",", "(", "2", ")", "where", "Mask", "Ddenotes", "a", "causal", "decoder", "mask", "And", "there", "is", "a", "projection", "layer", "for", "HN", "Dto", "get", "the", "final", "distribution", "of", "the", "predicted", "lyrics", "Training", "with", "Cycle", "-", "consistency", "Loss", "The", "main", "training", "loss", "is", "to", "minimize", "the", "negative", "log-", "likelihood", "over", "the", "lyrics", "L=", "(", "li", ")", "n", "i=1given", "the", "BeatMasked", "LyricLyricInput", "-", "to", "-", "Template", "ModuleDecoderEncoderLanguage", "ModelBeat’Lyric", "-", "to", "-", "Beat", "ModelLyric’Template", "-", "to", "-", "Lyric", "Module", "Figure", "4", ":", "Illustration", "of", "the", "training", "loss", "in", "the", "Template-", "to", "-", "Lyric", "module", "As", "described", "in", "Section", "3.2", ",", "we", "simplify", "the", "inputs", "for", "Bar", "B", "Segment", "Sand", "Intro", "-", "position", "Pcan", "be", "constructed", "with", "Masked", "Lyric", "M", ",", "Lyric", "L", ",", "and", "Beat", "B", "The", "Lyric", "-", "to", "-", "Beat", "model", "is", "a", "pre", "-", "trained", "model", "and", "details", "can", "be", "found", "in", "Appendix", "A", "template", "T=", "(", "ti", ")", "n", "i=1as", "shown", "in", "the", "gray", "dotted", "line", "in", "Figure", "4", ":", "LT2L=−logP", "(", "L|", "L", ",", "T", ")", "=", "−Σn", "i=1logP", "(", "li|", "l", "<", "i", ";", "t1", ",", "...", ",", "tn", ")", ",", "(", "3", ")", "where", "L=", "(", "li", ")", "n", "i=1denote", "the", "compound", "template", "The", "l", "<", "istands", "for", "sequence", "(", "l1", ",", "l2", ",", "...", ",", "li−1", ")", "As", "illustrated", "by", "the", "orange", "dotted", "line", "in", "Figure", "4", ",", "we", "introduce", "the", "cycle", "-", "consistency", "loss", "(", "CCL", ")", "to", "enhance", "the", "impact", "of", "musical", "information", "The", "Lyric", "-", "to", "-", "Beat", "model", "reconstructs", "the", "beat", "sequence", "from", "the", "predicted", "lyrics", "of", "the", "language", "model", "The", "formulation", "of", "CCL", "is", "as", "follows", ":", "L′", "L2B=−logP", "(", "B|L′", ")", "=", "−Σn", "i=1logP", "(", "bi|b", "<", "i", ";", "l′", "1", ",", "...", ",", "l′", "n", ")", ",", "(", "4", ")", "where", "L′=", "(", "l′", "i", ")", "n", "i=1denotes", "the", "predicted", "lyrics", "by", "the", "language", "model", ",", "and", "B=", "(", "bi", ")", "n", "i=1denotes", "the", "Beat", "of", "input", "template", "Tas", "in", "Figure", "4", "Finally", ",", "the", "training", "objective", "of", "Template", "-", "to-", "Lyric", "model", "is", "to", "minimize", "the", "loss", "Ltot", ":", "Ltot", "=", "LT2L+α∗", "L′", "L2B", ",", "(", "5", ")", "where", "num", "attention", "heads", "a", "hyper", "-", "parameter", "to", "weigh", "CCL" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "O", "B-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "I-TaskName", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "B-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterValue", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterName", "I-HyperparameterName", "I-HyperparameterName", "O", "O", "O", "O", "O", "O", "O" ]
[ "3.4", "Inference", "Procedure", "In", "this", "subsection", ",", "we", "describe", "the", "inference", "proce-", "dure", "of", "UniLG", "for", "various", "lyrics", "generation", "condi-", "tions", "The", "major", "steps", "are", "shown", "in", "Algorithm", "1,987", "including", "\"", "Beat", "Construction", "\"", ",", "\"", "Masked", "Lyric", "Con-", "struction", "\"", ",", "and", "\"", "Components", "Construction", "\"", "Given", "the", "template", "Tby", "Algorithm", "1", ",", "the", "Template", "-", "to-", "Lyric", "module", "generates", "the", "Lyric", "Land", "the", "Lauto-", "regressively", "\"", "Beat", "Construction", "\"", "is", "a", "method", "to", "construct", "the", "Beat", "Bfrom", "raw", "input", "X", "(", "e.g", ",", "beat", ",", "lyric", ",", "mu-", "sic", "score", ",", "audio", ",", "etc.", ")", "8", "\"", "Beat", "Construction", "\"", "con-", "sists", "of", "\"", "Sentence", "Length", "Generation", "\"", "and", "\"", "Beat", "Generation", "\"", "\"", "Sentence", "Length", "Generation", "\"", "gen-", "erates", "a", "sequence", "of", "numbers", "with", "each", "number", "denoting", "the", "length", "of", "one", "sentence9", "\"", "Beat", "Gen-", "eration", "\"", "generates", "the", "Beat", "based", "on", "the", "outputs", "from", "\"", "Sentence", "Length", "Generation", "\"", "For", "exam-", "ple", ",", "if", "\"", "Sentence", "Length", "Generation", "\"", "generate", "a", "sequence", "S=", "[", "3,2", "]", ",", "\"", "Beat", "Generation", "\"", "may", "re-", "turnB=", "[", "⟨bos⟩", ",", "Beat", "1", ",", "Beat", "3", ",", "Beat", "0", ",", "⟨", "/", "s⟩", ",", "Beat", "0", ",", "Beat", "1", ",", "⟨", "/", "s⟩", ",", "⟨eos⟩", "]", "To", "achieve", "content", "controllable", "generation", ",", "we", "use", "keywords", "Kto", "construct", "the", "Masked", "Lyrics", "M", "Based", "on", "the", "keyword10", ",", "the", "model", "generates", "the", "lyrics", "in", "the", "MLM", "manner", "The", "keywords", "can", "be", "either", "user", "-", "specified", "or", "sampled", "from", "the", "train-", "ing", "corpus", ",", "which", "should", "appear", "in", "the", "generated", "lyrics", "\"", "Masked", "Lyric", "Construction", "\"", "is", "a", "method", "to", "construct", "the", "masked", "lyrics", "Mcondition", "on", "Beat", "Band", "keywords", "K", "Similar", "to", "the", "construction", "of", "PandSin", "Section", "3.2", ",", "given", "Beat", "B", ",", "Mcan", "be", "constructed", "by", "replacing", "the", "token", "that", "is", "not", "in", "E", "with", "the", "mask", "token", "or", "keywords", "in", "Krandomly", "\"", "Components", "Construction", "\"", "is", "a", "method", "to", "ob-", "tain", "the", "other", "components", "given", "MandBas", "de-", "scribed", "in", "Section", "3.2", ",", "and", "organize", "all", "components", "for", "the", "template", "4", "Experimental", "Settings", "4.1", "Dataset", "We", "collect", "lyrics", "of", "249,007", "Chinese", "pop", "songs", "from", "Internet", "as", "the", "base", "of", "our", "experiments", "Lyric", "-", "Template", "Dataset", "We", "use", "the", "pre", "-", "trained", "Lyric", "-", "to", "-", "Beat", "model", "to", "extract", "the", "lyric", "-", "template", "dataset", "from", "249,007", "lyrics", "We", "randomly", "select", "8000", "songs", "for", "the", "validation", "and", "test", "set", "respec-", "tively", ",", "and", "the", "remaining", "songs", "are", "used", "for", "training", "The", "data", "statistics", "are", "shown", "in", "Appendix", "B", "8The", "details", "of", "the", "Lyric", "-", "to", "-", "Beat", ",", "MIDI", "-", "to", "-", "Beat", ",", "and", "Audio", "-", "to", "-", "Beat", "modules", "are", "discussed", "in", "Appendix", "D", "9The", "sentence", "means", "the", "sub", "-", "sequence", "between", "special", "symbols", "10If", "keywords", "are", "empty", ",", "we", "will", "randomly", "select", "some", "popular", "words", "as", "keywords", "Algorithm", "1", "Template", "Construction", "In", "Inference", "Input", ":", "X", ":", "the", "raw", "input", ";", "K", ":", "keywords", "Output", ":", "T", ",", "generated", "compound", "template", "DefBeat", "Construction", "(", "X", ")", ":", "Case", "of", "X", ":", "a", "beat", "sequence", ":", "B", "=", "X", "a", "lyric", "sequence", ":", "B", "=", "Lyric", "-", "to", "-", "Beat", "(", "X", ")", "a", "MIDI", "file", ":", "B", "=", "MIDI", "-", "to", "-", "Beat", "(", "X", ")", "a", "audio", "file", ":", "B", "=", "Audio", "-", "to", "-", "Beat", "(", "X", ")", "None", ":", "S", "=", "Sentence", "Length", "Generation", "(", ")", "B", "=", "Beat", "Generation", "(", "S", ")", "end", "Case", "return", "B", "B", "=", "Beat", "Construction", "(", "X", ")", "M", "=", "Masked", "Lyric", "Construction", "(", "B", ",", "K", ")", "return", "T", "=", "Components", "Construction", "(", "B", ",", "M", ")", "Additional", "Dataset", ":", "Song8k", "We", "also", "annotate", "8,000", "songs", "with", "structure", "labels", "(", "sentence", "-", "level", "chorus", "and", "verse", "label", ")", "for", "evaluation", "and", "we", "name", "this", "dataset", "Song8k", "For", "dataset", "settings", ",", "we", "use", "all", "8,000", "songs", "for", "further", "evaluation", "in", "the", "Template-", "to", "-", "Lyric", "module", "4.2", "Baselines", "in", "Model", "Comparison", "We", "compare", "with", "two", "baselines", "in", "the", "experi-", "ments", ":", "1", ")", "MT5", ",", "a", "pre", "-", "trained", "Transformer", "lan-", "guage", "model", "(", "Xue", "et", "al", ",", "2021b", ")", ";", "2", ")", "SongNet", ",", "a", "format", "-", "controlled", "text", "generation", "model", "(", "Li", "et", "al", ",", "2020", ")", "MT5", "and", "SongNet", "construct", "their", "inputs", "with", "the", "same", "corpus", "as", "the", "lyric", "-", "template", "dataset", "MT5", ",", "SongNet", ",", "and", "UniLG", "have", "similar", "parame-", "ters", "and", "all", "models", "use", "the", "same", "pre", "-", "trained", "model", "as", "initialization", "for", "a", "fair", "comparison", "The", "details", "of", "the", "model", "configuration", ",", "training", ",", "and", "decoding", "settings", "are", "reported", "in", "Appendix", "E", "and", "G", "4.3", "Objective", "Evaluation", "Metrics", "We", "use", "three", "kinds", "of", "objective", "evaluation", "metrics", ":", "general", "level", ",", "low", "level", ",", "and", "high", "level", "(", "more", "de-", "tails", "can", "be", "found", "in", "Appendix", "F", ")", "General", "Level", ":", "Besides", "accuracy", "(", "PPL", ")", ",", "we", "use", "RMSE", "to", "evaluate", "the", "sentence", "integrity", "(", "Li", "et", "al" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "B-MethodName", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ ",", "2020", ")", ",", "which", "calculates", "the", "average", "proba-", "bility", "of", "the", "separation", "token", "given", "previous", "tokens", "Low", "Level", ":", "We", "use", "perplexity", "and", "MAE", "to", "evaluate", "the", "degree", "of", "consistency", "between", "the", "gen-", "erated", "lyrics", "and", "the", "given", "textual", "format", "(", "Segment", "and", "Intro", "-", "Position", ")", "and", "rhythm", "patterns", "(", "Beat", ")", "in", "the", "template.988", "High", "Level", ":", "We", "use", "Song8k", "and", "a", "pre", "-", "trained", "model", "(", "details", "are", "in", "Appendix", "C", ")", "to", "evaluate", "the", "quality", "of", "the", "structure", "of", "generated", "lyrics" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "B-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Specifi-", "cally", ",", "the", "model", "predicts", "a", "chorus", "or", "verse", "label", "for", "each", "sentence", "in", "generated", "lyrics", "and", "compares", "it", "with", "the", "human", "annotations", "to", "obtain", "model_size" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName" ]
[ "4.4", "Subjective", "Evaluation", "Metrics", "As", "illustrated", "in", "Section", "?", "?", ",", "the", "Beat", "is", "important", "for", "the", "compound", "template", "and", "may", "have", "a", "big", "im-", "pact", "on", "our", "framework", "We", "conduct", "subjective", "ex-", "periments", "for", "\"", "Beat", "Construction", "\"", ",", "including", "\"", "Sen-", "tence", "Length", "Generation", "\"", "and", "\"", "Beat", "Generation", "\"", "Besides", ",", "we", "also", "conduct", "subjective", "experiments", "for", "model", "comparison", "For", "each", "subjective", "experi-", "ment", ",", "we", "invite", "43", "annotators", "to", "evaluate", "the", "gen-", "erated", "lyrics", "Each", "annotator", "is", "required", "to", "score", "lyrics", "concerning", "four", "aspects", "Each", "aspect", "is", "rated", "with", "an", "Spearman", "from", "1", "to", "5", "(", "from", "bad", "to", "excellent", ")" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "B-MetricValue", "O", "B-MetricValue", "O", "O", "O", "O", "O", "O" ]
[ "The", "four", "aspects", "are", "as", "follows", ":", "1", ")", "Co-", "herence", ":", "the", "overall", "consistency", "of", "the", "topic", "of", "the", "entire", "song", ";", "2", ")", "Fluency", ":", "the", "fluency", "of", "the", "semantic", "correlation", "within", "a", "sentence", "and", "between", "the", "sen-", "tence", ";", "3", ")", "Correlation", ":", "the", "structural", "or", "semantic", "similarity", "among", "sentences", ",", "such", "as", "the", "distribution", "of", "words", "and", "corresponding", "relationships", "of", "sen-", "tences", ";", "4", ")", "Fascination", ":", "the", "degree", "of", "fascinating", "sentences", "in", "annotators", "’", "opinion", "5", "Experiments", "Results", "In", "this", "section", ",", "we", "report", "and", "analyze", "both", "objective", "and", "subjective", "experimental", "results", "We", "also", "show", "some", "cases", "in", "Appendix", "H", "to", "verify", "the", "ability", "of", "UniLG", "to", "handle", "different", "generation", "conditions", "5.1", "Objective", "Results", "Model", "Comparison", "We", "compare", "MT5", ",", "SongNet", ",", "and", "UniLG", "on", "the", "Song8k", "and", "the", "test", "set", "of", "the", "lyric", "-", "template", "dataset", "The", "results", "of", "the", "model", "comparison", "(", "in", "Table", "1", ")", "show", "that", "MT5", "achieves", "the", "best", "results", "in", "model_size", "and", "MSE" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "B-MethodName", "O", "O", "B-MethodName", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "B-MetricName", "O", "B-MetricName" ]
[ "Our", "UniLG", "outperforms", "baselines", "in", "recall", ",", "RMSE", ",", "and", "time_to_converge", "log_likelihood", "shows", "that", "our", "framework", "does", "generate", "the", "lyrics", "with", "better", "structure", ",", "which", "indicates", "that", "the", "musical", "information", "improves", "the", "structural", "modeling" ]
[ "O", "B-MethodName", "O", "O", "O", "B-MetricName", "O", "B-MetricName", "O", "O", "B-MetricName", "B-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Ablation", "Study", "We", "further", "ablate", "our", "UniLG", "to", "verify", "the", "impact", "of", "musical", "and", "textual", "informa-", "tion", "as", "well", "as", "the", "CCL", "From", "the", "results", "shown", "in", "Table", "2", ",", "we", "see", "that", "the", "textual", "information", "(", "Seg", "&", "Pos", ")", ",", "musical", "information", "(", "Bar", "&", "Beat", ")", ",", "andCCL", "play", "crucial", "roles", "in", "the", "overall", "performance", "These", "modules", "of", "our", "framework", "show", "significant", "improvement", ",", "especially", "on", "the", "metrics", "of", "time_to_converge", "and", "ROUGE" ]
[ "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "B-MetricName" ]
[ "The", "CCL", "may", "enhance", "the", "musical", "information", "to", "boost", "performance", "in", "AUC", ",", "recall", ",", "and", "log_likelihood", ",", "but", "at", "the", "same", "time", "may", "introduce", "noise", "and", "cause", "degradation", "in", "general", "metrics", "(", "PPL", "and", "Integrity", ")" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "B-MetricName", "O", "O", "B-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "The", "effec-", "tiveness", "of", "the", "CCL", "further", "proves", "that", "the", "musi-", "cal", "information", "behind", "the", "lyrics", "does", "benefit", "the", "structure", "-", "aware", "lyric", "generation", "We", "notice", "that", "the", "musical", "information", "(", "Bar", "&", "Beat", ")", "degrades", "the", "per-", "formance", "of", "the", "framework", "more", "than", "the", "textual", "information", "and", "CCL", "This", "may", "be", "due", "to", "there", "being", "extra", "position", "embeddings", "for", "input", "data", "in", "MT5", "model", "When", "it", "comes", "to", "music", "information", ",", "missing", "Bar", "&", "Beat", "leads", "to", "a", "complete", "loss", "of", "infor-", "mation", ",", "while", "missing", "the", "Seg", "&", "Pos", "only", "partially", "loses", "position", "information", "5.2", "Subjective", "Results", "Template", "Construction", "As", "the", "template", "directly", "affects", "the", "Template", "-", "to", "-", "Lyric", "module", ",", "we", "perform", "the", "subjective", "evaluation", "on", "different", "settings", "of", "\"", "Sentence", "Length", "Generation", "\"", "and", "\"", "Beat", "Genera-", "tion", "\"", "in", "Algorithm", "1", "to", "investigate", "the", "impact", "of", "the", "compound", "template", "For", "\"", "Sentence", "Length", "Generation", "\"", ",", "we", "have", "2", "candidate", "settings", ":", "1", ")", "Random", ",", "the", "length", "of", "the", "sentence", "is", "randomly", "chosen", "from", "6", "to", "12", ";", "2", ")", "2-", "gram", ",", "the", "next", "sentence", "length", "is", "generated", "condi-", "tioned", "on", "the", "length", "of", "the", "previous", "sentence", "We", "generate", "40", "songs", "in", "6", "to", "16", "activation", "for", "each", "setting" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterValue", "O", "B-HyperparameterValue", "B-HyperparameterName", "O", "O", "O" ]
[ "Given", "two", "number", "sequences", "generated", "by", "\"", "Sentence", "Length", "Generation", "\"", "of", "two", "settings", ",", "the", "two", "Beat", "can", "be", "generated", "by", "the", "same", "method", "of", "\"", "Beat", "Generation", "\"", ",", "whose", "setting", "is", "chosen", "ran-", "domly", "The", "results", "in", "Table", "3", "show", "that", "Random", "and", "2", "-", "gram", "strategies", "achieve", "comparable", "perfor-", "mance", "and", "different", "sentence", "length", "generation", "strategies", "have", "little", "influence", "on", "models", "For", "\"", "Beat", "Generation", "\"", ",", "we", "have", "3", "candidate", "set-", "tings", ":", "1", ")", "Random", ",", "the", "beat", "information", "for", "each", "character", "is", "randomly", "chosen", "from", "B", ";", "2", ")", "Rule", ",", "the", "beat", "is", "non", "-", "decreasing", "than", "the", "previous", "one", ";", "3", ")", "Sample", ",", "we", "compute", "the", "statistics", "of", "the", "beat", "se-", "quence", "of", "each", "length", "in", "the", "lyric", "-", "template", "dataset", "and", "sample", "the", "beat", "sequence", "conditioned", "on", "the", "sequence", "length", "We", "generate", "40", "songs", "in", "6", "to", "16", "sentences", "for", "each", "setting", "(", "\"", "Sentence", "Lengths", "Gen-", "eration", "\"", "uses", "2", "-", "gram", ")", "The", "result", "in", "Table", "4", "shows989", "Table", "1", ":", "Model", "comparison", "with", "objective", "metrics", "on", "the", "test", "set", "of", "lyric", "-", "template", "dataset", "and", "Song8k", "T", "-", "L", "and", "S8", "stand", "for", "lyric", "-", "template", "dataset", "and", "Song8k", "respectively", "CCL", "denotes", "the", "cycle", "-", "consistency", "loss", "in", "Section", "?", "?", "Table", "3", ":", "The", "time_to_converge", "of", "different", "settings", "in", "\"", "Sen-", "tence", "Length", "Generation", "\"" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "I-DatasetName", "I-DatasetName", "O", "B-DatasetName", "B-DatasetName", "I-DatasetName", "I-DatasetName", "O", "B-DatasetName", "O", "O", "B-DatasetName", "I-DatasetName", "I-DatasetName", "I-DatasetName", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "We", "adapt", "2", "-", "gram", "for", "\"", "Sentence", "Length", "Generation", "\"", "and", "Sample", "for", "Beat", "Genera-", "tion", "in", "model", "comparison", "The", "results", "in", "Table", "5", "show", "that", "our", "UniLG", "outperforms", "the", "baselines", ",", "which", "furtherTable", "5", ":", "The", "Spearman", "of", "model", "comparison" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "O", "O" ]
[ "Table", "1", "and", "5", "prove", "that", "our", "compound", "template", "enables", "a", "stronger", "structure", "modeling", "ability", "than", "SongNet", "This", "may", "be", "because", "that", "the", "compound", "template", "provides", "discriminative", "representations", "for", "lyrics", "under", "the", "guidance", "of", "musical", "information", "The", "MT5", "achieves", "better", "Kendall", "and", "perplexity", "in", "Table", "1", "but", "gets", "lower", "model_size", "results", "in", "Fluency", "in", "Table", "5" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "B-MetricName", "O", "B-MetricName", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "O", "O", "O", "O", "O" ]
[ "This", "indicates", "that", "MT5", "may", "pay", "too", "much", "atten-", "tion", "to", "the", "fluency", "of", "the", "text", "but", "lacking", "the", "logical", "correlation", "between", "sentences", "5.3", "Case", "Studies", "We", "also", "show", "some", "cases", "in", "Appendix", "H", "to", "verify", "the", "ability", "of", "UniLG", "to", "handle", "different", "genera-", "tion", "conditions", "Although", "lyrics", "generated", "condi-", "tioned", "on", "the", "templates", "constructed", "by", "automatic", "method", "Template", "Construction", "are", "less", "satisfying", "(", "cases", "in", "Figure", "5", ")", ",", "the", "handcrafted", "template", "or", "template", "extracted", "by", "other", "resources", "achieves", "sat-", "isfying", "results", "as", "shown", "in", "Figure", "6", "and", "7", "These", "cases", "demonstrate", "that", "the", "template", "is", "human-990", "understandable", "and", "can", "be", "manipulated", "by", "users", "directly", "as", "in", "Section", "3.1", "The", "results", "in", "Figure", "8", "and", "9", "indicate", "that", "the", "template", "acts", "as", "a", "bridge", "between", "lyrics", "and", "the", "rhythmic", "sources", "(", "e.g.", ",", "au-", "dio", ",", "music", "score", ",", "etc", ")", ",", "which", "enables", "our", "UniLG", "to", "generate", "lyrics", "conditioned", "on", "different", "signals", "6", "Conclusion", "In", "this", "paper", ",", "we", "propose", "UniLG", ",", "a", "unified", "structure", "-", "aware", "lyric", "generation", "framework", "With", "our", "designed", "compound", "template", "to", "indicate", "the", "structure", "of", "lyrics", "with", "textual", "and", "musical", "infor-", "mation", ",", "which", "acts", "as", "a", "bridge", "between", "the", "rhyth-", "mic", "sources", "and", "lyrics", ",", "UniLG", "can", "handle", "different", "lyrics", "generation", "conditions", "We", "also", "introduce", "a", "cycle", "-", "consistency", "loss", "to", "enhance", "the", "impact", "of", "musical", "information", "to", "improve", "performance", "Ex-", "tensive", "experiments", "demonstrate", "the", "effectiveness", "of", "our", "framework", ",", "achieving", "significant", "improve-", "ment", "in", "both", "objective", "and", "subjective", "evaluations", "We", "will", "explore", "topic", "-", "driven", "lyrics", "generation", "in", "our", "future", "work", "Limitations", "The", "limitations", "of", "our", "work", "include", ":", "1", ")", "In", "our", "work", ",", "the", "structure", "of", "lyrics", "is", "the", "chorus", "and", "verse", "parts", "of", "songs", ",", "and", "it", "is", "learned", "in", "a", "data", "-", "driven", "man-", "ner", ",", "which", "highly", "relies", "on", "data", "quality", "2", ")", "The", "settings", "of", "the", "Lyric", "-", "to", "-", "Beat", "model", "will", "limit", "the", "effect", "of", "our", "model", "For", "this", "work", ",", "we", "make", "an", "assumption", "that", "all", "songs", "are", "with", "4", "/", "4", "time", "sig-", "natures", "for", "the", "Lyric", "-", "to", "-", "Beat", "model", "If", "the", "time", "signature", "is", "not", "4", "/", "4", ",", "we", "need", "to", "re", "-", "train", "the", "Lyric-", "to", "-", "Beat", "model.3", ")", "Better", "”", "Beat", "Construction", "”", "can", "be", "investigated", ",", "such", "as", "using", "a", "language", "model", "to", "generate", "the", "beat", "sequence", "We", "only", "explore", "the", "simple", "method", "and", "achieve", "satisfying", "results", "4", ")", "The", "model", "trained", "from", "scratch", "may", "not", "achieve", "satisfying", "results", "And", "a", "GPU", "with", "at", "least", "20", "G", "memory", "may", "be", "needed", "to", "use", "the", "pre", "-", "trained", "lan-", "guage", "model", "(", "MT5", ")", "to", "reproduce", "our", "work", "Ethics", "Statement", "Under", "the", "review", "of", "the", "company", "’s", "legal", "team", ",", "the", "data", "collected", "for", "research", "is", "under", "legally", "cor-", "rect", "copyright", "The", "artifacts", "we", "used", "(", "e.g.", ",", "MT5", ",", "other", "codes", ",", "etc", ")", "are", "consistent", "with", "their", "intended", "use", "and", "meet", "corresponding", "licenses", "The", "mother", "tongue", "of", "all", "annotators", "is", "Chinese", "and", "the", "annota-", "tors", "are", "recruited", "by", "the", "human", "resources", "depart-", "ments", "and", "the", "payment", "is", "adequate", "enough", "(", "The", "an", "-", "notators", "receive", "an", "hourly", "wage", "of", "80", "RMB", ",", "about", "12", "USD", ")", "according", "to", "the", "laws", "and", "regulations", "of", "our", "country", "Before", "the", "experiments", ",", "we", "report", "key", "information", "about", "the", "requirements", "for", "human", "an-", "notators", ",", "including", "the", "evaluation", "criteria", "and", "the", "usage", "of", "their", "annotations", "We", "have", "used", "the", "data", "under", "the", "consensus", "of", "the", "industry", "and", "research", "and", "the", "final", "information", "used", "for", "research", "does", "not", "include", "any", "protected", "category" ]
[ "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Proceedings", "of", "the", "61st", "Annual", "Meeting", "of", "the", "Association", "for", "Computational", "Linguistics", "Volume", "1", ":", "Long", "Papers", ",", "pages", "1002–1017", "July", "9", "-", "14", ",", "2023", "©", "2023", "Association", "for", "Computational", "Linguistics", "FC", "-", "KBQA", ":", "A", "Fine", "-", "to", "-", "Coarse", "Composition", "Framework", "for", "Knowledge", "Base", "Question", "Answering", "Lingxi", "Zhang1", ",", "Jing", "Zhang1∗", ",", "Yanling", "Wang1", ",", "Shulin", "Cao2", ",", "Xinmei", "Huang1", ",", "Cuiping", "Li1", ",", "Hong", "Chen1", ",", "Juanzi", "Li2", "1School", "of", "Information", ",", "Renmin", "University", "of", "China", ",", "Beijing", ",", "China", "2Department", "of", "Computer", "Science", "and", "Technology", ",", "Tsinghua", "University", ",", "Beijing", ",", "China", "{", "zhanglingxi", ",", "zhang", "-", "jing", ",", "wangyanling", ",", "huangxinmei", ",", "licuiping", ",", "chong", "}", "@", "ruc.edu.cn", "{", "caosl19", "}", "@", "mails.tsinghua.edu.cn", ",", "{", "lijuanzi", "}", "@", "tsinghua.edu.cn", "Abstract", "The", "generalization", "problem", "on", "KBQA", "has", "drawn", "considerable", "attention", "Existing", "research", "suffers", "from", "the", "generalization", "issue", "brought", "by", "the", "entanglement", "in", "the", "coarse", "-", "grained", "mod-", "eling", "of", "the", "logical", "expression", ",", "or", "inexecutabil-", "ity", "issues", "due", "to", "the", "fine", "-", "grained", "modeling", "of", "disconnected", "classes", "and", "relations", "in", "real", "KBs", "We", "propose", "a", "Fine", "-", "to", "-", "Coarse", "Compo-", "sition", "framework", "for", "KBQA", "(", "FC", "-", "KBQA", ")", "to", "both", "ensure", "the", "generalization", "ability", "and", "exe-", "cutability", "of", "the", "logical", "expression", "The", "main", "idea", "of", "FC", "-", "KBQA", "is", "to", "extract", "relevant", "fine-", "grained", "knowledge", "components", "from", "KB", "and", "reformulate", "them", "into", "middle", "-", "grained", "knowl-", "edge", "pairs", "for", "generating", "the", "final", "logical", "ex-", "pressions", "FC", "-", "KBQA", "derives", "new", "state", "-", "of", "-", "the-", "art", "performance", "on", "GrailQA", "and", "WebQSP", ",", "and", "runs", "4", "times", "faster", "than", "the", "baseline", "Our", "code", "is", "now", "available", "at", "GitHub", "https", ":", "/", "/", "github", "1", "Introduction", "Question", "answering", "over", "knowledge", "bases", "(", "KBQA", ")", "aims", "to", "provide", "a", "user", "-", "friendly", "way", "to", "access", "large-", "scale", "knowledge", "bases", "(", "KBs", ")", "by", "natural", "language", "questions", "Existing", "KBQA", "methods", "(", "Zhang", "et", "al", ",", "2023", ")", "can", "be", "roughly", "categorized", "into", "retrieval-", "based", "and", "semantic", "-", "parsing", "(", "SP", ")", "based", "methods", "The", "former", "(", "Feng", "et", "al", ",", "2021", ";", "He", "et", "al", ",", "2021a", ";", "Zhang", "et", "al", ",", "2022", ")", "directly", "scores", "the", "relevance", "between", "the", "question", "and", "answer", "candidates", ",", "thus", "it", "is", "difficult", "to", "resolve", "the", "complex", "questions", "On", "the", "contrary", ",", "some", "KBQA", "approaches", ",", "such", "as", "(", "Das", "et", "al", ",", "2021", ";", "Kapanipathi", "et", "al", ",", "2021", ";", "Qiu", "et", "al", ",", "2020", ";", "Sun", "et", "al", ",", "2020", ")", ",", "are", "based", "on", "semantic", "pars-", "ing", "(", "denoted", "as", "SP", "-", "based", ")", ",", "which", "can", "address", "com-", "plex", "questions", "and", "achieve", "promising", "results", "on", "i.i.d", "datasets", "SP", "-", "based", "methods", "first", "translate", "the", "questions", "into", "logical", "expressions", "such", "as", "SPARQL", "and", "then", "execute", "them", "against", "KB", "to", "yield", "answers", "∗Corresponding", "author", "Figure", "1", ":", "Illustration", "of", "generalization", "tasks", "in", "KBQA", "Each", "question", "is", "paired", "with", "a", "logical", "expression", "that", "consists", "of", "different", "components", "Components", "involved", "in", "the", "training", "data", "are", "colored", "in", "non", "-", "green", "color", ",", "while", "unseen", "components", "are", "colored", "in", "green", "The", "coarse", "-", "grained", "method", "directly", "matches", "the", "question", "with", "the", "logical", "expression", "(", "i.e.", ",", "the", "composition", "of", "components", ")", ",", "while", "the", "fine", "-", "grained", "method", "matches", "the", "question", "with", "each", "component", "candidate", "and", "then", "composes", "them", "to", "derive", "the", "logical", "expression", "The", "MAP", "of", "logical", "expressions", "on", "compositional", "generalization", "test", "data", "and", "recall", "test", "data", "is", "shown", "on", "the", "right", "of", "the", "figure" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "O", "B-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "As", "illustrated", "in", "Figure", "1", ",", "a", "logical", "expression", "con-", "sists", "of", "multiple", "components", "such", "as", "classes", "and", "relations", "Most", "existing", "SP", "-", "based", "approaches", "fail", "with", "logical", "expressions", "that", "contain", "unseen", "compo-", "sitions", "of", "components", "(", "called", "Kendall", ")", "or", "unseen", "components", "(", "called", "recall", ")" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "O", "O", "O", "O", "O", "B-MetricName", "O" ]
[ "To", "address", "the", "above", "problem", ",", "GrailQA-", "Rank", "(", "Gu", "et", "al", ",", "2021", ")", "proposes", "a", "BERT", "-", "based", "rank", "model", "to", "match", "the", "given", "question", "with", "each", "logical", "expression", "candidate", ",", "which", "leverages", "the", "generalization", "abilities", "of", "the", "pre", "-", "trained", "language", "models", "On", "top", "of", "that", ",", "RNG", "-", "KBQA", "(", "Ye", "et", "al", ",1002", "2022", ")", "further", "uses", "a", "pre", "-", "trained", "generation", "model", ",", "which", "takes", "top-5", "ranked", "logical", "expressions", "as", "the", "additional", "input", "beyond", "the", "question", "to", "generate", "the", "target", "logical", "expression", "Behind", "these", "main-", "stream", "models", ",", "a", "logical", "expression", "is", "viewed", "as", "an", "inseparable", "unit", "during", "modeling", "Actually", ",", "logical", "expressions", "are", "coarse", "-", "grained", "because", "they", "can", "be", "decomposed", "into", "relatively", "fine", "-", "grained", "com-", "ponents", "including", "relations", ",", "classes", ",", "entities", ",", "and", "logical", "skeletons", "(", "See", "examples", "in", "Figure", "3", ")", "Such", "coarse", "-", "grained", "modeling", "entangles", "representations", "of", "fine", "-", "grained", "components", ",", "thereby", "overfitting", "the", "seen", "compositions", "during", "the", "training", "process", ",", "which", "weakens", "the", "model", "’s", "compositional", "general-", "ization", "ability", "Meanwhile", ",", "even", "though", "pre", "-", "trained", "language", "models", "can", "deal", "with", "zero", "-", "shot", "compo-", "nents", "to", "some", "extent", ",", "compositional", "overfit", "reduces", "their", "ability", "to", "identify", "individual", "unseen", "compo-", "nents", "with", "zero", "-", "shot", "generalization", "To", "demonstrate", "the", "above", "idea", ",", "we", "perform", "a", "pilot", "study", "(", "Cf", "the", "detailed", "settings", "in", "Section", "4.1", ")", "with", "two", "preliminary", "experiments", ":", "one", "calculates", "the", "similarity", "score", "between", "a", "question", "and", "each", "coarse", "-", "grained", "logical", "expression", "to", "obtain", "the", "most", "relevant", "one", ",", "and", "the", "other", "searches", "the", "most", "relevant", "fine", "-", "grained", "components", "to", "form", "the", "final", "logical", "expression", "of", "a", "question", "We", "observe", "that", "the", "fine", "-", "grained", "modeling", "derives", "more", "accurate", "logical", "expressions", "on", "both", "the", "compositional", "task", "and", "zero", "-", "shot", "task", "(", "Cf", "It", "could", "be", "explained", "that", "fine", "-", "grained", "modeling", "focuses", "ex-", "clusively", "on", "each", "component", ",", "avoiding", "overfitting", "of", "seen", "compositions", "in", "the", "training", "data", "Although", "some", "studies", "attempt", "to", "leverage", "fine", "-", "grained", "com-", "ponents", ",", "they", "only", "consider", "partial", "fine", "-", "grained", "components", "such", "as", "relations", ",", "classes", ",", "and", "entities", "(", "Chen", "et", "al", ",", "2021", ")", ",", "or", "suffer", "from", "inexecutability", "due", "to", "disconnected", "fine", "-", "grained", "components", "in", "real", "KBs", "(", "Shu", "et", "al", ",", "2022", ")", "Thus", ",", "to", "both", "ensure", "the", "generalization", "ability", "and", "executability", "of", "logical", "expressions", ",", "we", "pro-", "pose", "a", "Fine", "-", "to-", "Coarse", "composition", "framework", "for", "KBQA", "(", "FC", "-", "KBQA", ")", ",", "which", "contains", "three", "sub-", "modules", "The", "overview", "of", "our", "model", "is", "shown", "in", "Figure", "4", "The", "first", "module", "is", "fine", "-", "grained", "com-", "ponent", "detection", ",", "which", "detects", "all", "kinds", "of", "fine-", "grained", "component", "candidates", "from", "Freebase", "by", "their", "semantic", "similarities", "with", "the", "question", "Such", "component", "detection", "guarantees", "the", "generalization", "ability", "in", "both", "compositional", "and", "zero", "-", "shot", "tasks", "The", "second", "module", "is", "the", "middle", "-", "grained", "compo", "-", "nent", "constraint", ",", "which", "efficiently", "prunes", "and", "com-", "poses", "the", "fine", "-", "grained", "component", "candidates", "by", "en-", "suring", "the", "components", "’", "connectivity", "in", "the", "KB", "The", "final", "module", "is", "the", "coarse", "-", "grained", "component", "com-", "position", ",", "which", "employs", "a", "seq", "-", "to", "-", "seq", "generation", "model", "to", "generate", "the", "executable", "coarse", "-", "grained", "logical", "expression", "In", "addition", "to", "encode", "the", "fine-", "grained", "components", ",", "the", "middle", "-", "grained", "compo-", "nents", "are", "also", "encoded", "to", "enhance", "the", "model", "’s", "rea-", "soning", "capacity", ",", "so", "as", "to", "improve", "the", "executability", "of", "the", "generated", "logical", "expression", "In", "contrast", "to", "previous", "work", "(", "Cao", "et", "al", ",", "2022b", ";", "Chen", "et", "al", ",", "2021", ";", "Shu", "et", "al", ",", "2022", ")", "that", "only", "uses", "the", "knowl-", "edge", "constraints", "to", "guide", "the", "decoding", "process", ",", "we", "emphasize", "injecting", "them", "into", "the", "encoding", "pro-", "cess", ",", "because", "the", "encoder", "which", "learns", "bidirec-", "tional", "context", "could", "better", "suit", "natural", "language", "understanding", "(", "Du", "et", "al", ",", "2022", ")", "We", "conduct", "extensive", "experiments", "on", "widely", "used", "GrailQA", ",", "WebQSP", ",", "and", "CWQ", "datasets", "GrailQA", "(", "Gu", "et", "al", ",", "2021", ")", "is", "a", "KBQA", "bench-", "mark", "focusing", "on", "generalization", "problems", "FC-", "KBQA", "derives", "new", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "GrailQA", "-", "Dev", "(", "+7.6", "%", "perplexity", "gain", "and", "+7.0", "%", "recall", "gain", "respectively", ")" ]
[ "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "O", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "B-DatasetName", "O", "O", "B-DatasetName", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "B-MetricValue", "I-MetricValue", "B-MetricName", "O", "O", "B-MetricValue", "I-MetricValue", "B-MetricName", "O", "O", "O" ]
[ "Meanwhile", ",", "FC", "-", "KBQA", "also", "obtains", "good", "performance", "on", "WebQSP", "and", "CWQ", "Moreover", ",", "FC", "-", "KBQA", "runs", "4", "times", "faster", "than", "the", "state", "-", "of", "-", "the", "-", "art", "baseline", "RNG", "-", "KBQA", "The", "abla-", "tion", "studies", "demonstrate", "the", "effect", "of", "our", "middle-", "grained", "encoding", "strategy", "Contributions", "(", "1", ")", "We", "conduct", "a", "pilot", "study", "to", "reveal", "an", "intriguing", "phenomenon", "—", "a", "fine", "-", "grained", "understanding", "of", "the", "logical", "expression", "helps", "en-", "hance", "the", "generalization", "ability", "of", "SP", "-", "based", "KBQA", "methods", ",", "which", "is", "rarely", "discussed", "before", "(", "2", ")", "We", "propose", "a", "fine", "-", "to", "-", "coarse", "composition", "framework", "FC", "-", "KBQA", "to", "address", "the", "generalization", "problem", ",", "which", "takes", "advantage", "of", "the", "idea", "of", "fine", "-", "grained", "modeling", "(", "3", ")", "We", "devise", "a", "middle", "-", "grained", "compo-", "nent", "constraint", "that", "is", "injected", "into", "both", "the", "encoder", "and", "the", "decoder", "to", "guide", "the", "seq", "-", "to", "-", "seq", "model", "in", "producing", "executable", "logical", "expressions", "(", "4", ")", "FC", "-", "KBQA", "not", "only", "maintains", "efficiency", "but", "also", "achieves", "significant", "improvement", "on", "GrailQA", "2", "Related", "Work", "Coarse", "-", "Grained", "SP", "-", "based", "Methods", "Many", "ef-", "forts", "are", "paid", "to", "solve", "generalization", "problems", "on", "SP", "-", "based", "KBQA", "Some", "approaches", ",", "such", "as", "(", "Lan", "and", "Jiang", ",", "2020", ";", "Gu", "et", "al", ",", "2021", ")", ",", "use", "a", "rank", "-", "based", "model", "that", "takes", "advantage", "of", "a", "coarse", "-", "level", "match1003", "between", "the", "question", "and", "the", "logical", "expressions", "or", "query", "graphs", "They", "first", "enumerate", "numerous", "query", "graph", "candidates", "based", "on", "KBs", "and", "then", "they", "rank", "them", "according", "to", "how", "relevant", "they", "are", "to", "the", "ques-", "tion", "Another", "line", "of", "approaches", ",", "in", "addition", "to", "the", "rank", "-", "based", "ones", ",", "makes", "use", "of", "a", "generation", "model", "KQAPro", "(", "Cao", "et", "al", ",", "2022a", ")", "leverages", "BART", "to", "directly", "convert", "questions", "into", "logical", "expressions", "Additionally", ",", "RNG", "-", "KBQA", "(", "Ye", "et", "al", ",", "2022", ")", "further", "injects", "top", "-", "k", "ranked", "logical", "expressions", "as", "an", "ad-", "ditional", "input", "to", "the", "question", "CBR", "-", "KBQA", "(", "Das", "et", "al", ",", "2021", ")", "injects", "analogous", "questions", "and", "their", "corresponding", "logical", "expressions", "from", "the", "training", "data", "to", "increase", "the", "generalization", "All", "of", "the", "afore-", "mentioned", "methods", "are", "pure", "coarse", "-", "level", "frame-", "works", "that", "treat", "each", "coarse", "-", "grained", "logical", "expres-", "sion", "as", "a", "separate", "unit", "Fine", "-", "Grained", "SP", "-", "based", "Methods", "Many", "re-", "searchers", "have", "been", "motivated", "to", "address", "the", "gen-", "eralization", "issue", "by", "the", "notion", "of", "utilizing", "decom-", "posed", "components", ",", "such", "as", "class", ",", "relation", ",", "and", "log-", "ical", "skeleton", "Some", "approaches", "(", "Wang", "et", "al", ",", "2020", ";", "Zhao", "et", "al", ",", "2022", ";", "Li", "et", "al", ",", "2023", ")", "retrieve", "the", "rele-", "vant", "schema", "item", "such", "as", "relation", "and", "column", "as", "additional", "fine", "-", "grained", "input", "information", ",", "while", "an-", "other", "line", "of", "approaches", "(", "Dong", "and", "Lapata", ",", "2018", ")", "extracts", "the", "skeleton", "of", "logical", "expression", "as", "the", "de-", "coder", "guide", "Such", "methods", "primarily", "concentrate", "on", "the", "grammar", "of", "logical", "expression", "and", "often", "ignore", "the", "knowledge", "constraint", ",", "which", "is", "essen-", "tial", "in", "large", "-", "scale", "KB", "They", "usually", "focus", "on", "KBs", "or", "DBs", "that", "contain", "a", "small", "number", "of", "relations", "where", "a", "logical", "expression", "can", "be", "easy", "to", "be", "exe-", "cutable", "Program", "Transfer", "(", "Cao", "et", "al", ",", "2022b", ")", ",", "Re-", "Track", "(", "Chen", "et", "al", ",", "2021", ")", ",", "and", "TIARA", "(", "Shu", "et", "al", ",", "2022", ")", "simply", "apply", "KB", "constraints", "to", "control", "the", "generation", "of", "the", "decoding", "process", "As", "opposed", "to", "them", ",", "we", "make", "use", "of", "middle", "-", "grained", "KB", "con-", "straints", "during", "both", "the", "encoding", "and", "the", "decoding", "processes", "to", "help", "the", "model", "better", "adapt", "to", "KB", "and", "ensure", "executability", "3", "Problem", "Definition", "Knowledge", "Base", "(", "KB", ")", "A", "KB", "is", "comprised", "by", "ontology", "{", "(", "C×R×C", ")", "}", "and", "relational", "facts", "{", "(", "E×R×", "(", "E∪C", ")", ")", "}", ",", "where", "R", ",", "C", ",", "andEdenote", "relation", "set", ",", "class", "set", ",", "and", "entity", "set", "respectively", "Notably", ",", "we", "consider", "literal", "as", "a", "special", "type", "of", "en-", "tity", "Specifically", ",", "an", "ontology", "triple", "(", "cd", ",", "r", ",", "cr", ")", "con-", "sists", "of", "a", "relation", "r∈R", ",", "a", "domain", "class", "cdwhich", "denotes", "the", "class", "of", "the", "subject", "entities", ",", "and", "a", "rangeclasscrwhich", "denotes", "the", "class", "of", "the", "object", "enti-", "ties", "Each", "class", "has", "multiple", "entities", ",", "thus", "an", "ontol-", "ogy", "triplet", "can", "be", "instantiated", "as", "several", "relational", "facts", "For", "example", ",", "both", "(", "e1", ",", "r", ",", "e", "2", ")", "and", "(", "e3", ",", "r", ",", "e", "4", ")", "correspond", "to", "(", "cd", ",", "r", ",", "cr", ")", ",", "where", "e1", ",", "e3∈cdand", "e2", ",", "e4∈cr", "Figure", "3", "illustrates", "a", "KB", "subgraph", "SP", "-", "based", "KBQA", "Given", "a", "natural", "question", "q", ",", "KBQA", "models", "aim", "to", "find", "a", "set", "of", "entities", "denoted", "byA⊆Efrom", "KB", "as", "the", "answers", "to", "q", "Instead", "of", "directly", "predicting", "A", ",", "SP", "-", "based", "KBQA", "models", "translate", "qto", "an", "executable", "logical", "expression", "de-", "noted", "by", "ssuch", "as", "SPARQL", ",", "lambda", "-", "DCS", "(", "Liang", "et", "al", ",", "2013", ")", ",", "query", "graph", "(", "Lan", "and", "Jiang", ",", "2020", ")", ",", "and", "s", "-", "expression", "(", "Gu", "et", "al", ",", "2021", ")", "We", "select", "s", "-", "expression", "as", "our", "used", "logical", "ex-", "pression", "since", "it", "could", "provide", "a", "good", "trade", "-", "off", "on", "compactness", ",", "compositionality", ",", "and", "readabil-", "ity", "(", "Gu", "et", "al", ",", "2021", ")", "The", "logical", "skeleton", "of", "an", "s", "-", "expression", "can", "be", "derived", "by", "removing", "all", "the", "relations", ",", "classes", ",", "and", "entities", "in", "the", "expression", "and", "only", "keeping", "function", "operators", "and", "parenthe-", "ses", "Specifically", ",", "we", "replace", "relations", ",", "classes", ",", "enti-", "ties", ",", "literals", "with", "special", "tokens", "“", "<", "rel", ">", "”", ",", "“", "<", "class", ">", "”", ",", "“", "<", "entity", ">", "”", ",", "“", "<", "literal", ">", "”", "respectively", "Figure", "3", "shows", "an", "executable", "logical", "expression", "on", "the", "KB", "and", "its", "corresponding", "logical", "skeleton", "We", "unitedly", "name", "the", "relations", ",", "classes", ",", "entities", ",", "and", "logical", "skeleton", "in", "an", "s", "-", "expression", "as", "the", "fine", "-", "grained", "component", ",", "while", "the", "complete", "s", "-", "expression", "is", "thecoarse", "-", "grained", "logical", "expression", "4", "Approach", "4.1", "Pilot", "Study", "As", "analyzed", "in", "Section", "1", ",", "considering", "the", "logical", "expression", "as", "a", "unit", "will", "lead", "to", "entangled", "represen-", "tations", "of", "fine", "-", "grained", "components", "and", "thus", "weak-", "ens", "generalization", "ability", "Here", "we", "study", "the", "ne-", "cessity", "of", "fine", "-", "grained", "modeling", "by", "testing", "how", "coarse", "-", "grained", "and", "fine", "-", "grained", "matching", "methods", "perform", "when", "selecting", "a", "question", "’s", "logical", "expres-", "sion", "from", "the", "corresponding", "candidate", "pool", "Dataset", "To", "simplify", "the", "experiment", ",", "we", "extract", "a", "toy", "dataset", "that", "only", "involves", "1", "-", "hop", "logical", "expres-", "sions", "from", "GrailQA", "Then", ",", "for", "the", "relation", "rand", "the", "class", "cin", "such", "logical", "expressions", ",", "we", "study", "the", "compositional", "generalization", "where", "the", "compo-", "sition", "(", "r", ",", "c", ")", "is", "unseen", "or", "zero", "-", "shot", "generalization", "where", "the", "individual", "rorcis", "unseen", "in", "the", "training", "data", "For", "each", "question", "with", "its", "ground", "-", "truth", "logi-", "cal", "expression", ",", "we", "select", "100", "logical", "expressions1004", "Figure", "3", ":", "Illustration", "of", "a", "KB", "subgraph", "and", "an", "exe-", "cutable", "logical", "expression", ",", "where", "the", "ovals", "denote", "the", "entities", ",", "the", "rectangles", "denote", "the", "classes", ",", "the", "solid", "lines", "denote", "the", "relations", ",", "and", "the", "dashed", "lines", "connect", "the", "entities", "and", "their", "classes", "The", "upper", "part", "of", "the", "subgraph", "illustrates", "examples", "of", "ontology", "triplets", ",", "while", "the", "bot-", "tom", "illustrates", "relational", "facts", "that", "share", "the", "same", "domain", "as", "the", "ground", "truth", "as", "the", "coarse", "-", "grained", "expression", "candidates", "For", "fair", "comparison", ",", "we", "separate", "all", "of", "the", "relations", ",", "classes", ",", "and", "logical", "skeletons", "from", "the", "coarse", "-", "grained", "can-", "didates", "as", "the", "fine", "-", "grained", "component", "candidates", "Methods", "We", "aim", "to", "find", "the", "target", "logical", "expres-", "sion", "of", "a", "given", "question", "by", "a", "ranking", "model", "trained", "with", "a", "contrastive", "loss", "(", "Chen", "et", "al", ",", "2020", ")", ",", "which", "is", "also", "used", "by", "RNG", "-", "KBQA", "(", "Ye", "et", "al", ",", "2022", ")", "The", "coarse", "-", "grained", "method", "concatenates", "a", "question", "and", "a", "candidate", "logical", "expression", "to", "feed", "into", "BERT", ",", "then", "the", "output", "embedding", "of", "[", "CLS", "]", "is", "fed", "into", "a", "linear", "layer", "to", "compute", "the", "similarity", "score", "The", "fine", "-", "grained", "method", "follows", "the", "above", "pipeline", ",", "but", "the", "input", "is", "the", "concatenation", "of", "a", "question", "and", "a", "fine", "-", "grained", "candidate", "component", ",", "then", "scores", "each", "logical", "expression", "candidate", "by", "summing", "up", "the", "normalized", "question", "-", "component", "similarity", "scores", "For", "both", "methods", ",", "we", "compute", "accuracy", "by", "evalu-", "ating", "whether", "the", "ground", "-", "truth", "logical", "expression", "owns", "the", "highest", "score", "in", "the", "candidate", "pool", "Observation", "—", "Fine", "-", "grained", "modeling", "can", "bet-", "ter", "solve", "the", "generalization", "problems", "on", "KBQA", "The", "cross_entropy", "is", "reported", "in", "Figure", "2" ]
[ "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "B-DatasetName", "O", "B-DatasetName", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "B-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "O", "B-MetricName", "O", "O", "O", "O", "O" ]
[ "The", "fine", "-", "grained", "method", "outperforms", "the", "coarse-", "grained", "method", "in", "both", "composition", "generaliza-", "tion", "and", "zero", "-", "shot", "generalization", "tasks", "A", "possi-", "ble", "explanation", "is", "the", "fine", "-", "grained", "matching", "fo", "-", "cuses", "solely", "on", "each", "component", "and", "is", "simple", "to", "learn", ",", "which", "better", "capture", "the", "semantic", "informa-", "tion", "of", "each", "component", "and", "also", "well", "adaptable", "to", "express", "the", "various", "compositions", "of", "components", "The", "coarse", "-", "grained", "matching", ",", "on", "the", "other", "hand", ",", "at-", "tempts", "to", "describe", "all", "of", "the", "components", "as", "a", "whole", "composition", ",", "limiting", "the", "ability", "to", "express", "unseen", "compositions", "and", "components", "Inspired", "by", "this", ",", "we", "propose", "FC", "-", "KBQA", "in", "the", "next", "section", "4.2", "Model", "Overview", "We", "propose", "a", "fine", "-", "to", "-", "coarse", "composition", "frame-", "work", "FC", "-", "KBQA", "bridged", "by", "a", "middle", "-", "grained", "KB", "constraint", "Figure", "4", "illustrates", "the", "overall", "frame-", "work", ",", "which", "contains", "three", "parts", ":", "Fine", "-", "grained", "Component", "Detection", "Given", "a", "question", ",", "we", "extract", "relation", "candidates", "and", "class", "candidates", "from", "the", "whole", "KB", "based", "on", "seman-", "tic", "similarity", "Simultaneously", ",", "we", "adopt", "an", "entity", "linker", "to", "detect", "mentioned", "entities", "and", "use", "a", "seq", "-", "to-", "seq", "model", "to", "generate", "logical", "skeletons", "Middle", "-", "grained", "Component", "Constraint", "Based", "on", "the", "detected", "components", ",", "we", "devise", "an", "efficient", "way", "to", "check", "the", "connectivity", "of", "component", "pairs", "on", "the", "KB", ",", "including", "class", "-", "relation", "pairs", ",", "relation-", "relation", "pairs", ",", "and", "relation", "-", "entity", "pairs", "We", "only", "keep", "the", "executable", "component", "pairs", "to", "guarantee", "the", "executability", "of", "final", "logical", "expression", "Coarse", "-", "grained", "Component", "Composition", "Fi-", "nally", ",", "a", "seq", "-", "to", "-", "seq", "model", "takes", "the", "concatenation", "of", "the", "question", "and", "the", "reformulated", "components", "as", "input", "to", "generate", "the", "logical", "expression", "In", "par-", "ticular", ",", "the", "middel", "-", "grained", "components", "are", "injected", "into", "both", "the", "encoder", "and", "the", "decoder", "to", "ensure", "the", "executability", "of", "the", "final", "logical", "expressions", "4.3", "Fine", "-", "grained", "Component", "Detection", "Relation", "and", "Class", "Extraction", "Taking", "the", "re-", "lation", "extractor", "as", "the", "example", ",", "given", "a", "question", "q", ",", "we", "aim", "to", "extract", "relations", "in", "q", "First", ",", "we", "apply", "BM25", "(", "Robertson", "et", "al", ",", "2009", ")", "to", "recall", "the", "relation", "candidates", "from", "the", "KB", "based", "on", "the", "surface", "over-", "laps", "between", "relations", "’", "names", "and", "q", "Then", "we", "ap-", "ply", "BERT", "(", "Devlin", "et", "al", ",", "2019", ")", "as", "the", "cross", "-", "encoder", "to", "measure", "the", "semantic", "similarity", "between", "qand", "each", "relation", "candidate", "r", "We", "describe", "rusing", "the", "relation", "domain", ",", "the", "relation", "name", ",", "and", "the", "rela-", "tion", "range", "and", "let", "the", "BERT", "input", "be", "“", "[", "CLS", "]", "q", "[", "D", "]", "domain", "(", "r", ")", "[", "N", "]", "name", "(", "r", ")", "[", "R", "]", "range", "(", "r", ")", "[", "SEP", "]", "”", ",", "where", "[", "CLS", "]", ",", "[", "SEP", "]", ",", "[", "D", "]", ",", "[", "N", "]", ",", "and", "[", "R", "]", "are", "the", "special", "to-", "kens", "To", "better", "distinguish", "the", "spurious", "relations,1005", "Figure", "4", ":", "Overview", "of", "FC", "-", "KBQA", "In", "the", "step", "of", "fine", "-", "grained", "component", "detection", ",", "we", "perform", "class", "extraction", ",", "relation", "extraction", ",", "entity", "linking", ",", "and", "logical", "skeleton", "parsing", "to", "obtain", "the", "most", "relevant", "components", "of", "the", "question", "Then", "we", "utilize", "the", "KB", "-", "based", "constraint", "to", "obtain", "middle", "-", "grained", "component", "pairs", "that", "are", "connected", "in", "the", "KB", "Finally", ",", "a", "T5", "-", "based", "seq", "-", "to", "-", "seq", "model", "encodes", "the", "reformulated", "fine", "-", "grained", "and", "middle", "-", "grained", "candidates", "(", "reformulation", "unit", ")", ",", "and", "employs", "a", "controllable", "decoder", "with", "dynamic", "vocabulary", "(", "control", "unit", ")", "to", "generate", "the", "executable", "target", "logical", "expression", "we", "sample", "the", "relations", "that", "share", "the", "same", "domain", "as", "the", "ground", "-", "truth", "relation", "as", "the", "negatives", "for", "training", "The", "trained", "model", "is", "used", "to", "retrieve", "the", "set", "of", "top-", "krelations", ",", "denoted", "by", "Rq", "The", "class", "extractor", "works", "in", "the", "same", "way", "as", "the", "relation", "extractor", "We", "represent", "the", "class", "using", "its", "name", "and", "domain", ",", "and", "use", "other", "classes", "in", "the", "same", "domain", "as", "negatives", "Cqrepresents", "the", "set", "of", "the", "top-", "krelevant", "classes", "Entity", "Linking", "A", "common", "paradigm", "of", "finding", "topic", "entities", "in", "KBQA", "methods", "is", "to", "first", "leverage", "a", "NER", "tool", "(", "Finkel", "et", "al", ",", "2005", ")", "to", "detect", "mentions", "and", "then", "apply", "an", "entity", "disambiguation", "model", "to", "link", "them", "to", "entities", "in", "KB", "However", ",", "some", "noun-", "phrase", "mentions", "such", "as", "“", "rich", "media", "”", "are", "hard", "to", "be", "detected", "by", "the", "NER", "tool", ",", "and", "some", "ambiguous", "entities", "could", "not", "be", "distinguished", "by", "the", "pure", "en-", "tity", "names", "To", "address", "both", "issues", ",", "we", "equip", "the", "NER", "tool1with", "a", "trie", "tree", "-", "based", "mention", "detec-", "tion", "method", "and", "propose", "a", "relation", "-", "aware", "pruning", "method", "to", "filter", "the", "mentions", "Specifically", ",", "we", "build", "a", "trie", "tree", "(", "Fredkin", ",", "1960", ")", "with", "the", "surface", "names", "of", "all", "entities", "in", "the", "KB", "Then", "we", "can", "search", "noun", "phrase", "mentions", "in", "the", "question", "efficiently", "and", "link", "them", "to", "the", "KB", "by", "1We", "follow", "GrailQA", "which", "utilizes", "an", "open", "BERT", "-", "NER", "tool", "on", "GitHub", "(", "https", ":", "/", "/", "github.com", "/", "kamalkraj", "/", "BERT-NER", ")", ".BLINK", "(", "Wu", "et", "al", ",", "2020", ")", "to", "obtain", "the", "correspond-", "ing", "entities", "Eq", "After", "that", ",", "we", "propose", "a", "relation", "awared", "pruning", "strategy", "to", "prune", "Eqby", "removing", "the", "entities", "that", "could", "not", "link", "to", "any", "relations", "in", "Rq", "Finally", ",", "following", "GrailQA", "(", "Gu", "et", "al", ",", "2021", ")", ",", "we", "choose", "the", "entity", "with", "the", "highest", "popularity", "We", "define", "regular", "expressions", "to", "extract", "literals", "such", "as", "digits", "and", "years", "appearing", "in", "q", "Logical", "Skeleton", "Parsing", "Logical", "skeleton", "pars-", "ing", "aims", "to", "transform", "a", "given", "question", "qinto", "a", "logical", "skeleton", "l", "Because", "the", "logical", "skeleton", "is", "domain", "-", "independent", ",", "the", "parsing", "process", "could", "be", "generalized", "across", "domains", "We", "adopt", "T5", "(", "Raffel", "et", "al", ",", "2020", ")", ",", "a", "state", "-", "of", "-", "the", "-", "art", "generation", "model", "to", "parse", "logical", "skeletons", "Since", "many", "entity", "names", "contain", "tokens", "such", "as", "“", "and", "”", "and", "“", "of", "”", "that", "may", "cause", "the", "logical", "skeleton", "to", "be", "incorrectly", "deter-", "mined", ",", "we", "mask", "each", "mention", "m∈Mqwith", "the", "special", "token", "“", "<", "entity0", ">", "”", ",", "“", "<", "entity1", ">", "”", ",", "...", ",", "in", "order", "of", "appearance", "For", "example", ",", "we", "change", "“", "Thomas", "was", "the", "designer", "of", "what", "ship", "?", "”", "to", "“", "<", "entity0", ">", "was", "the", "designer", "of", "what", "ship", "?", "”", "We", "notice", "that", "a", "common", "error", "is", "parsing", "out", "logical", "skeleton", "with", "wrong", "relation", "numbers", ",", "for", "example", "“", "<", "rel", ">", "”", "in-", "stead", "of", "“", "<", "rel", ">", "<", "rel", ">", "”", "Instead", "of", "increasing", "beam", "numbers", ",", "we", "manually", "add", "grammar", "rules", ",", "such", "as", "add", "“", "<", "rel", ">", "<", "rel", ">", "”", "as", "the", "second", "candidate", "when1006", "“", "<", "rel", ">", "”", "is", "T5", "’s", "top-1", "prediction", "The", "set", "of", "the", "top-2", "logical", "skeleton", "candidates", "is", "denoted", "as", "Lq", "4.4", "Middle", "-", "grained", "Component", "Constrain", "After", "deriving", "the", "candidate", "components", "according", "to", "Section", "4.3", ",", "the", "KB", "-", "based", "constraint", "is", "required", "to", "guarantee", "the", "composed", "logical", "expression", "is", "executable", "A", "straightforward", "idea", "is", "to", "fill", "the", "logical", "skeleton", "with", "candidate", "relations", ",", "classes", ",", "and", "entities", ",", "and", "execute", "them", "one", "by", "one", "to", "check", "executability", "However", ",", "such", "enumeration", "is", "ineffi-", "cient", ",", "since", "all", "combinations", "of", "candidate", "compo-", "nents", "should", "be", "considered", "Therefore", ",", "we", "incor-", "porate", "the", "middle", "-", "grained", "component", "pairs", "which", "are", "connected", "in", "KB", "Such", "pairs", "can", "be", "produced", "efficiently", "to", "keep", "the", "model", "’s", "efficiency", "The", "middle", "-", "grained", "component", "pairs", "include", "class", "-", "relation", "pairs", ",", "relation", "-", "relation", "pairs", ",", "and", "relation", "-", "entity", "pairs", "For", "each", "class", "c∈Cqand", "each", "relation", "r∈Rq", ",", "ifris", "connected", "with", "the", "domain", "class", "c", ",", "we", "add", "(", "c", ",", "r", ")", "into", "the", "class-", "relation", "pair", "set", "Pc−r", "For", "example", "in", "Figure", "3", ",", "the", "class", "“", "railway.railway", "”", "is", "linked", "with", "the", "re-", "lation", "“", "rail.railway.terminuses", "”", ",", "so", "the", "pair", "(", "rail-", "way.railway", ",", "rail.railway.terminuses", ")", "is", "executable", "and", "will", "be", "added", "into", "Pc−r", "If", "the", "range", "class", "ofrisc", ",", "we", "add", "the", "pair", "of", "cand", "the", "reverse", "relation", "of", "r", "We", "construct", "executable", "relation-", "relation", "pair", "set", "Pr−rby", "checking", "each", "relation", "pair", "(", "r1∈Rq", ",", "r2∈Rq", ")", "Ifr2", "’s", "domain", "class", "does", "not", "match", "r1", "’s", "range", "class", ",", "we", "directly", "re-", "move", "this", "pair", "to", "maintain", "efficiency", ",", "otherwise", ",", "we", "reformulate", "(", "r1", ",", "r2", ")", "to", "a", "logical", "expression", "and", "execute", "on", "KB", "to", "check", "its", "connectivity", "For", "each", "relation", "-", "entity", "pair", "(", "r", ",", "e", ")", ",", "we", "first", "check", "whether", "the", "logical", "skeleton", "candidates", "contain", "the", "<", "entity", ">", "placeholder", "or", "not", "If", "not", ",", "we", "leave", "Pr−eempty", ";", "otherwise", "we", "directly", "take", "the", "result", "of", "the", "relation-", "pruning", "strategy", "for", "entities", "in", "Section", "4.3", "4.5", "Coarse", "-", "grained", "Component", "Composition", "We", "apply", "a", "generation", "model", "based", "on", "T5", "to", "com-", "pose", "all", "the", "above", "fine", "-", "grained", "and", "middle", "-", "grained", "component", "candidates", "and", "output", "an", "executable", "logical", "expression", "by", "a", "controlled", "decoder", "Encoding", "Process", "Before", "feeding", "the", "fine-", "grained", "and", "middle", "-", "grained", "component", "candidates", "into", "the", "generator", ",", "we", "sort", "the", "middle", "-", "grained", "can-", "didates", "according", "to", "their", "similarity", "scores", "to", "the", "question", "By", "doing", "this", ",", "the", "order", "can", "reveal", "the", "pattern", "of", "which", "pair", "is", "more", "likely", "to", "appear", "in", "theground", "-", "truth", "logical", "expression", "In", "intuition", ",", "such", "a", "pattern", "will", "help", "to", "generate", "more", "accurate", "log-", "ical", "expressions", "To", "accomplish", "this", ",", "we", "take", "the", "logits", "of", "the", "fine", "-", "grained", "component", "detection", "in", "section", "4.3", "as", "the", "similarity", "score", "between", "the", "ques-", "tion", "and", "each", "class", "/", "relation", "component", ",", "and", "then", "calculate", "the", "similarity", "score", "between", "the", "question", "and", "a", "middle", "-", "grained", "component", "pair", "by", "summing", "the", "scores", "of", "contained", "single", "components", "The", "encoding", "of", "such", "middle", "-", "grained", "component", "im-", "proves", "the", "generator", "’s", "reasoning", "capacity", "in", "terms", "of", "capturing", "the", "knowledge", "constraints", "We", "use", "“", ";", "”", "to", "separate", "each", "element", "(", "a", "compo-", "nent", "or", "a", "component", "pair", ")", "To", "explicitly", "inform", "the", "model", "the", "type", "of", "each", "component", ",", "we", "place", "“", "[", "REL", "]", "”", ",", "“", "[", "CL", "]", "”", ",", "“", "[", "ENT", "]", "”", ",", "and", "“", "[", "LF", "]", "”", "before", "each", "relation", ",", "class", ",", "entity", ",", "and", "logical", "skeleton", "respectively", "For", "example", ",", "we", "organize", "the", "input", "of", "encoder", "as", "“", "query", ";", "[", "CL", "]", "c1", "[", "REL", "]", "r1", ";", "[", "REL", "]", "r1", "[", "REL", "]", "r2", ";", "[", "CL", "]", "c2", "[", "REL", "]", "r3", ";", "[", "ENT", "]", "e1", ";", "[", "LF", "]", "l1", ";", "[", "LF", "]", "l2", "”", "Decoding", "Process", "The", "middle", "-", "grained", "compo-", "nents", "are", "also", "used", "to", "produce", "a", "dynamic", "vocab-", "ulary", "to", "constrain", "the", "decoding", "process", "The", "gen-", "erated", "token", "ytis", "confined", "to", "the", "tokens", "involved", "in", "the", "dynamic", "vocabulary", "at", "each", "step", "t", "We", "ini-", "tialize", "the", "dynamic", "vocabulary", "with", "the", "union", "of", "tokens", "from", "the", "detected", "entities", ",", "tokens", "from", "the", "detected", "classes", "in", "Pc−r", ",", "i.e.", ",", "usually", "the", "answer", "type", ",", "and", "the", "keywords", "such", "as", "“", "JOIN", "”", "in", "logical", "skeleton", "Then", "we", "update", "the", "dynamic", "vocabulary", "by", "the", "relations", "paired", "with", "rinPr−rif", "the", "last", "generated", "component", "is", "ror", "by", "the", "relations", "paired", "withcinPc−rif", "it", "is", "c", "5", "Experiment", "5.1", "Experimental", "Settings", "Dataset", "We", "evaluate", "our", "method", "on", "GrailQA", "(", "Gu", "et", "al", ",", "2021", ")", ",", "WebQSP", "(", "Yih", "et", "al", ",", "2016", ")", ",", "and", "CWQ", "(", "Talmor", "and", "Berant", ",", "2018", ")", ",", "all", "of", "which", "are", "based", "on", "Freebase", "GrailQA", "focuses", "on", "generaliza-", "tion", "problems", "which", "involved", "up", "to", "4", "-", "hop", "logical", "expressions", "and", "complex", "operations", "WebQSP", "is", "an", "i.i.d", "benchmark", "that", "required", "2", "-", "hop", "reasoning", "Although", "CWQ", "is", "not", "designed", "to", "solve", "generaliza-", "tion", "problem", ",", "we", "can", "still", "separate", "out", "the", "zero", "-", "shot", "test", "set", "with", "all", "the", "unseen", "relations", "and", "classes", ",", "yielding", "576", "/", "3519", "zero", "-", "shot", "/", "all", "test", "set", "Evaluation", "Metrics", "To", "measure", "the", "cross_entropy", "of", "logical", "expression", ",", "we", "use", "the", "F1_score", "which", "measures", "the", "exact", "equivalence1007", "Table", "1", ":", "Results", "of", "overall", "evaluation", "on", "GrailQA", "-", "LeaderBoard", "(", "%", ")" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O" ]
[ "Overall", "I.I.D", "We", "also", "calculate", "the", "MAP", "based", "on", "the", "predicted", "and", "gold", "answers" ]
[ "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "O", "O", "O", "O", "O", "O" ]
[ "Baselines", "On", "GrailQA", ",", "we", "mainly", "compare", "with", "the", "published", "works", "on", "the", "leaderboard", ",", "in-", "cluding", "GrailQA", "-", "Rank", "(", "Gu", "et", "al", ",", "2021", ")", ",", "GrailQA-", "Trans", "(", "Gu", "et", "al", ",", "2021", ")", ",", "Retrack", "(", "Chen", "et", "al", ",", "2021", ")", ",", "RNG", "-", "KBQA", "(", "Ye", "et", "al", ",", "2022", ")", "They", "are", "all", "SP-", "based", "models", "that", "target", "generalization", "problems", "in", "KBQA", "On", "WebQSP", "and", "CWQ", ",", "we", "compare", "our", "method", "with", "the", "retrieval", "-", "based", "models", "in-", "cluding", "GraphNet", "(", "Pu", "et", "al", ",", "2018", ")", ",", "PullNet", "(", "Sun", "et", "al", ",", "2019", ")", "and", "NSM", "(", "He", "et", "al", ",", "2021b", ")", ",", "and", "the", "SP", "-", "based", "models", "including", "QGG", "(", "Lan", "and", "Jiang", ",", "2020", ")", ",", "RNG", "-", "KBQA", "(", "Ye", "et", "al", ",", "2022", ")", ",", "and", "PI", "Trans-", "fer", "(", "Cao", "et", "al", ",", "2022b", ")", "We", "evaluate", "precision", "for", "the", "retrieval", "-", "based", "models", ",", "while", "evaluate", "both", "log_likelihood", "and", "MAP", "for", "the", "SP", "-", "based", "methods" ]
[ "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "O", "B-DatasetName", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "B-MetricName", "O", "O", "O", "O", "O", "O" ]
[ "We", "compare", "all", "the", "baselines", "that", "have", "the", "results", "on", "the", "two", "datasets", "or", "publish", "the", "codes", "that", "can", "be", "executed", "5.2", "Overall", "Evaluation", "Performance", "In", "Table", "1", "and", "Table", "2", ",", "we", "eval-", "uate", "the", "performance", "of", "FC", "-", "KBQA", "on", "different", "datasets", "For", "the", "baselines", ",", "we", "directly", "take", "their", "results", "reported", "in", "the", "original", "papers", "To", "be", "noted", ",", "on", "the", "extracted", "zero", "-", "shot", "test", "set", "of", "CWQ", ",", "the", "re-", "sults", "for", "some", "models", "remain", "empty", "because", "their", "full", "codes", "are", "not", "released", "As", "shown", "in", "Table", "1", ",", "our", "model", "outperforms", "all", "the", "baselines", ",", "especially", "on", "the", "compositional", "and", "zero", "-", "shot", "test", "tasks", "Com-", "pared", "with", "RNG", "-", "KBQA", ",", "the", "state", "-", "of", "-", "the", "-", "art", "pub-", "lished", "model", ",", "we", "have", "an", "absolute", "gain", "of", "4.3", "%", "and", "4.4", "%", "in", "terms", "of", "AUC", "and", "Spearman", "respectively" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricValue", "I-MetricValue", "O", "B-MetricValue", "I-MetricValue", "O", "O", "O", "B-MetricName", "O", "B-MetricName", "O" ]
[ "We", "also", "outperform", "on", "the", "extracted", "zero", "-", "shot", "CWQ", "test", "set", "by", "11.3", "%", "in", "terms", "of", "cross_entropy", ",", "as", "for", "an", "unseen", "complex", "question", ",", "parsing", "out", "correct", "knowledge", "components", "and", "logical", "skeletons", "is", "much", "easier", "than", "directly", "parsing", "the", "coarse", "-", "grained", "logical", "ex-", "pression", "correctly" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "B-MetricValue", "I-MetricValue", "O", "O", "O", "B-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Since", "the", "fine", "-", "grained", "module", "solely", "focuses", "on", "each", "component", "and", "thus", "leadsTable", "2", ":", "MAE", "Evaluation", "on", "WebQSP", "and", "CWQ", "(", "%", ")" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "O", "B-DatasetName", "O", "B-DatasetName", "O", "O", "O" ]
[ "On", "the", "original", "test", "set", "of", "CWQ", ",", "we", "only", "under", "-", "perform", "PI", "Transfer", "which", "leverages", "a", "pre", "-", "train", "process", "on", "a", "large", "-", "scale", "wiki", "data", "that", "is", "out", "scope", "of", "CWQ", "Efficiency", "Both", "RNG", "-", "KBQA", "and", "GrailQA", "-", "Rank", "enumerate", "all", "the", "logical", "expressions", "in", "a", "2", "-", "hop", "KB", "subgraph", "(", "enumeration", ")", ",", "so", "it", "is", "time", "-", "consuming", "for", "the", "rank", "model", "to", "score", "thousands", "of", "logical", "expressions", "for", "each", "question", "(", "candidate", "selection", ")", "Conversely", ",", "our", "FC", "-", "KBQA", "just", "retrieves", "the", "most", "relevant", "components", "(", "candidate", "selection", ")", "and", "then", "enumerates", "the", "component", "pairs", "based", "on", "the", "fil-", "tered", "candidates", "(", "enumeration", ")", ",", "which", "greatly", "re-", "duces", "the", "inference", "time", "Besides", "enumeration", "and", "candidate", "selection", ",", "a", "seq", "-", "to", "-", "seq", "model", "is", "used", "to", "generate", "the", "final", "logical", "expression", "(", "final", "composi-", "tion", ")", "In", "the", "same", "24", "GB", "GPU", "and", "Intel", "Gold", "5218", "CPU", ",", "the", "experimental", "results", "in", "Figure", "5", "show", "that", "our", "model", "runs", "4", "times", "faster", "than", "baselines", "5.3", "Ablation", "Studies", "GrailQA", "does", "not", "provide", "ground", "truth", "for", "the", "test", "set", ",", "so", "we", "conduct", "the", "ablation", "studies", "on", "the", "public", "Grail", "-", "Dev", "to", "investigate", "how", "the", "fine-", "and", "middle-", "grained", "components", "affect", "the", "performance", "As", "shown", "in", "Table", "3", ",", "we", "develop", "four", "model", "variants", "(", "1", ")", "-Knowledge", "removes", "all", "the", "fine-", "grained", "and", "middle", "-", "grained", "components", "except", "for", "the", "logical", "skeleton", "(", "2", ")", "-Knowledge", "Pairs", "replaces", "the", "middle", "-", "grained", "components", ",", "such", "as1008", "Table", "3", ":", "Ablation", "studies", "on", "GrailQA", "-", "Dev", "(", "%", ")", "Overall", "I.I.D", "“", "Overall", "”", "de-", "notes", "the", "total", "inference", "time", "of", "each", "model", "Specially", ",", "GrailQA", "-", "Rank", "has", "no", "composition", "step", ",", "we", "record", "the", "corresponding", "time", "as", "zero", "class", "-", "relation", "pairs", "and", "relation", "-", "relation", "pairs", "with", "the", "corresponding", "fine", "-", "grained", "candidates", ",", "such", "as", "classes", "and", "relations", "(", "3", ")", "-Logical", "Skeleton", "gets", "rid", "of", "the", "logical", "skeleton", "(", "4", ")", "-Decode", "Constraint", "deletes", "the", "dynamic", "vocabulary", "created", "with", "the", "middle", "-", "grained", "components", "The", "results", "show", "that", "removing", "“", "knowledge", "”", "re-", "duces", "model", "performance", "by", "60", "%", "cross_entropy", ",", "and", "replacing", "“", "knowledge", "pairs", "”", "with", "pure", "fine", "-", "grained", "components", "also", "reduces", "model", "performance", "by", "28", "%", "accuracy", ",", "indicating", "that", "encoding", "the", "middle-", "grained", "components", "can", "significantly", "improve", "the", "model", "’s", "reasoning", "capacity" ]
[ "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricValue", "I-MetricValue", "B-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricValue", "I-MetricValue", "B-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "To", "further", "demonstrate", "that", "encoding", "such", "middle", "-", "grained", "components", "can", "also", "help", "improve", "other", "model", "’s", "performance", ",", "we", "create", "Enhanced", "RNG", "-", "KBQA", "by", "taking", "the", "top-10", "ranked", "results", "from", "its", "ranking", "model", "and", "formulat-", "ing", "them", "into", "middle", "-", "grained", "component", "pairs", "to", "be", "injected", "into", "its", "encoder", "The", "results", "in", "Table", "3", "show", "that", "middle", "-", "grained", "reformulation", "improves", "the", "performance", "of", "RNG", "-", "KBQA", "Middle", "-", "grained", "component", "pairs", ",", "like", "coarse", "-", "grained", "logical", "ex-", "pressions", ",", "can", "guarantee", "connectivity", ",", "but", "they", "aremore", "compact", "and", "much", "shorter", "As", "a", "result", ",", "be-", "cause", "PLMs", "have", "a", "maximum", "input", "length", ",", "the", "middle", "-", "grained", "formulation", "can", "inject", "more", "com-", "ponents", "and", "is", "more", "likely", "to", "cover", "the", "components", "involved", "in", "the", "target", "logical", "expression", "Removing", "“", "logical", "skeleton", "”", "can", "result", "in", "a", "3.0", "%", "MSE", "drop", ",", "indicating", "that", "skeleton", "is", "useful", "for", "guid-", "ing", "the", "question", "understanding", "even", "though", "it", "is", "less", "important", "than", "the", "knowledge" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricValue", "I-MetricValue", "B-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Removing", "“", "decode", "constraint", "”", "in", "the", "decoder", "can", "also", "have", "an", "effect", "on", "model", "performance", ",", "but", "is", "much", "weaker", "than", "removing", "“", "knowledge", "pairs", "”", "in", "the", "encoder", ",", "indicating", "that", "injecting", "the", "knowl-", "edge", "constraints", "in", "the", "encoding", "process", "is", "more", "useful", "than", "in", "the", "decoding", "process", ",", "because", "the", "en-", "coder", "learns", "the", "bidirectional", "context", ",", "which", "is", "bet-", "ter", "suited", "to", "natural", "language", "understanding", "This", "is", "also", "a", "significant", "difference", "from", "the", "existing", "knowledge", "constrained", "decoding", "methods", "Both", "\"", "Knowledge", "Pairs", "\"", "and", "\"", "Decode", "Con-", "straint", "\"", "are", "proposed", "for", "addressing", "the", "in-", "executability", "issue", ",", "which", "guarantee", "all", "generated", "logical", "expressions", "are", "executable", "Removing", "ei-", "ther", "reduces", "the", "accuracy", ",", "which", "indicates", "that", "high", "executability", "can", "improve", "the", "model", "performance", "5.4", "Error", "Analysis", "We", "randomly", "select", "50", "error", "cases", "on", "GrailQA", "and", "summarize", "the", "error", "into", "three", "main", "categories", ":", "error", "entity", "(", "60", "%", ")", ",", "error", "relation", "and", "class", "(", "35", "%", ")", ",", "and", "error", "logical", "skeleton", "(", "40", "%", ")", "We", "also", "analysis", "the", "error", "cases", "while", "our", "model", "fails", "but", "some", "baseline", "methods", "can", "answer", "successfully", "resolve", "them", "A", "typical", "mistake", "is", "on", "logical", "expressions", "that", "involve", "KB", "-", "specific", "component", "composition", "For", "example", ",", "in", "Freebase", ",", "“", "coach", "”", "is", "represented", "by", "the", "join", "of", "“", "sports.sports_team.coaches", "”", "and", "“", "sports.sports_team_coach_tenure.coach”.1009", "Our", "fine", "-", "to", "-", "coarse", "model", "only", "predicts", "the", "previous", "relation", "but", "is", "unable", "to", "recall", "“", "sports.sports_team_coach_tenure.coach", "”", ",", "while", "some", "coarse", "-", "grained", "methods", "are", "able", "to", "memorize", "such", "composition", "and", "provide", "the", "correct", "answer", "6", "Conclusion", "This", "paper", "proposes", "FC", "-", "KBQA", ",", "a", "Fine", "-", "to-", "Coarse", "composition", "framework", "for", "KBQA", "The", "core", "idea", "behind", "it", "is", "to", "solve", "the", "entanglement", "issue", "of", "main-", "stream", "coarse", "-", "grained", "modeling", "by", "the", "fine", "-", "grained", "modeling", ",", "and", "further", "improve", "the", "executability", "of", "logical", "expression", "by", "reformulating", "the", "fine-", "grained", "knowledge", "into", "middle", "-", "grained", "knowledge", "pairs", "Benefiting", "from", "this", ",", "FC", "-", "KBQA", "achieves", "new", "state", "-", "of", "-", "the", "-", "art", "performance", "and", "efficiency", "on", "the", "compositional", "and", "zero", "-", "shot", "generalization", "KBQA", "tasks", "This", "fine", "-", "to", "-", "coarse", "framework", "with", "middle", "-", "grained", "knowledge", "injection", "could", "be", "in-", "spiring", "for", "generalization", "on", "other", "NLP", "tasks", "7", "Limitations", "Although", "our", "model", "achieves", "good", "performance", "in", "solving", "the", "compositional", "and", "zero", "-", "shot", "generalization", "problems", ",", "there", "is", "still", "room", "for", "improvement", "on", "the", "i.i.d", "datasets", "The", "fine", "-", "grained", "module", "in", "our", "framework", "can", "not", "take", "advantage", "of", "explicit", "composition", "infor-", "mation", "when", "the", "component", "compositions", "in", "the", "testing", "set", "and", "training", "set", "significantly", "overlapp", "For", "example", ",", "in", "Freebase", ",", "\"", "Who", "is", "the", "coach", "of", "FC", "Barcelona", "?", "\"", "is", "answered", "by", "the", "join", "of", "relation", "“", "sports.sports_team.coaches", "”", "and", "“", "sports.sports_team_coach_tenure.coach", "”", "Our", "fine", "-", "grained", "extractor", "may", "fail", "to", "recall", "“", "sports.sports_team_coach_tenure.coach", "”", "and", "in-", "stead", "select", "“", "base.american_football.football_coac", "-h.coach", "”", "as", "the", "candidate", "since", "‘", "football", "coach", "”", "is", "more", "relevant", "to", "the", "question", "than", "“", "coach", "tenure", "”", "in", "semantics", "The", "only", "coarse", "-", "grained", "model", ",", "however", ",", "can", "directly", "memorize", "the", "pattern", "because", "such", "composition", "appears", "frequently", "in", "the", "training", "data", "Therefore", ",", "compared", "to", "conventional", "models", "that", "completely", "memorize", "composition", "patterns", ",", "our", "model", "may", "only", "have", "minor", "advantages", "Another", "limitation", "is", "that", "we", "can", "not", "guar-", "antee", "the", "generalization", "on", "other", "KBs", "such", "as", "WikiData", "because", "gaps", "between", "KBs", "may", "bring", "negative", "impact", "For", "example", ",", "rela-", "tions", "in", "Freebase", "are", "often", "more", "specific", "(", "ice_hockey.hockey_player.hockey_position", ",", "soccer.football_player.position_s", ")", ",", "while", "re-", "lations", "in", "Wikidata", "are", "more", "general", "(", "posi-", "tion_played_on_team", ")", "We", "consider", "it", "as", "a", "direction", "for", "our", "future", "work", "8", "Ethics", "Statement", "This", "work", "focuses", "on", "the", "generalization", "issue", "of", "knowledge", "base", "question", "answering", ",", "and", "the", "con-", "tribution", "is", "fully", "methodological", "Hence", ",", "there", "are", "no", "direct", "negative", "social", "impacts", "of", "this", "work", "For", "experiments", ",", "this", "work", "uses", "open", "datasets", "that", "have", "been", "widely", "used", "in", "previous", "work", "and", "are", "without", "sensitive", "information", "as", "we", "know", "The", "authors", "of", "this", "work", "follow", "the", "ACL", "Code", "of", "Ethics", "and", "the", "application", "of", "this", "work", "have", "no", "obvious", "issue", "that", "may", "lead", "to", "the", "risk", "of", "ethics" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Proceedings", "of", "the", "61st", "Annual", "Meeting", "of", "the", "Association", "for", "Computational", "Linguistics", "Volume", "1", ":", "Long", "Papers", ",", "pages", "1018–1032", "July", "9", "-", "14", ",", "2023", "©", "2023", "Association", "for", "Computational", "Linguistics", "Does", "GPT-3", "Grasp", "Metaphors", "?", "Identifying", "Metaphor", "Mappings", "with", "Generative", "Language", "ModelsLennart", "WachowiakKing", "’s", "College", "Londonlennart.wachowiak", "@", "gmail.comDagmar", "GromannUniversity", "of", "Viennadagmar.gromann", "@", "gmail.comAbstractConceptual", "metaphors", "present", "a", "powerful", "cog", "-", "nitive", "vehicle", "to", "transfer", "knowledge", "structuresfrom", "a", "source", "to", "a", "target", "domain", "Prior", "neuralapproaches", "focus", "on", "detecting", "whether", "natu", "-", "ral", "language", "sequences", "are", "metaphoric", "or", "lit", "-", "eral", "We", "believe", "that", "to", "truly", "probe", "metaphoricknowledge", "in", "pre", "-", "trained", "language", "models", ",", "their", "capability", "to", "detect", "this", "transfer", "should", "beinvestigated", "To", "this", "end", ",", "this", "paper", "proposes", "toprobe", "the", "ability", "of", "GPT-3", "to", "detect", "metaphoriclanguage", "and", "predict", "the", "metaphor", "’s", "source", "do", "-", "main", "without", "any", "pre", "-", "set", "domains", "We", "experi", "-", "ment", "with", "different", "training", "sample", "configura", "-", "tions", "for", "fine", "-", "tuning", "and", "few", "-", "shot", "prompting", "ontwo", "distinct", "datasets", "When", "provided", "12", "few", "-", "shot", "samples", "in", "the", "prompt", ",", "GPT-3", "generatesthe", "correct", "source", "domain", "for", "a", "new", "samplewith", "an", "recall", "of", "65.15", "%", "in", "English", "and34.65", "%", "in", "Spanish" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "B-TaskName", "I-TaskName", "O", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "B-MetricValue", "I-MetricValue", "O", "O", "B-MetricValue", "I-MetricValue", "O", "O" ]
[ "GPT", "’s", "most", "common", "er", "-", "ror", "is", "a", "hallucinated", "source", "domain", "for", "whichno", "indicator", "is", "present", "in", "the", "sentence", "Othercommon", "errors", "include", "identifying", "a", "sequenceas", "literal", "even", "though", "a", "metaphor", "is", "presentand", "predicting", "the", "wrong", "source", "domain", "basedon", "specific", "words", "in", "the", "sequence", "that", "are", "notmetaphorically", "related", "to", "the", "target", "domain.1", "IntroductionMetaphor", "processing", "with", "pre", "-", "trained", "languagemodels", "(", "e.g.", "Conneau", "et", "al", ",2020", ";", "Brown", "et", "al", ",2020", ")", "has", "been", "dominated", "by", "metaphor", "detec", "-", "tion", ",", "that", "is", ",", "the", "classification", "of", "expressions", "intometaphoric", "or", "literal", "(", "e.g.", "Aghazadeh", "et", "al", ",2022", ";", "Leong", "et", "al", ",2020", ")", "In", "metaphor", "interpretation", ",", "acommon", "approach", "is", "to", "paraphrase", "metaphoric", "ex", "-", "pressions", "into", "literal", "ones", "(", "e.g.", "Stowe", "et", "al", ",2021a", ")", ".Few", "approaches", "target", "metaphor", "identification", ",", "e.g.predicting", "the", "source", "domain", "of", "a", "metaphor", "in", "a", "lin", "-", "guistic", "sequence", "For", "instance", ",", "Rosen", "(", "2018", ")", "relieson", "grammatical", "constructs", "and", "pre", "-", "defined", "labels", "Instead", ",", "in", "this", "paper", ",", "we", "test", "a", "generative", "languagemodel", "’s", "ability", "to", "predict", "the", "source", "domain", "givena", "target", "domain", "and", "sequence", "without", "grammaticalassumptions", "or", "fixed", "source", "domain", "labels", "Conceptual", "metaphor", "theory", "(", "CMT", ")", "(", "Lakoff", "andJohnson,1980", ")", "starts", "from", "the", "assumption", "thatmetaphors", "represent", "a", "powerful", "cognitive", "mech", "-", "anism", "to", "transfer", "physical", "knowledge", "structures", "toabstract", "domains", "In", "natural", "language", ",", "He", "wasbombarded", "by", "insultsorYour", "words", "pierce", "myhearttransfers", "the", "concrete", "domain", "of", "weaponsto", "the", "abstract", "domain", "of", "words", "in", "the", "metaphorWORDS", "ARE", "WEAPONS", "On", "the", "assumption", "thatour", "cognitive", "organization", "relies", "on", "metaphors", ",", "au", "-", "tomatically", "identifying", "metaphoric", "transfer", "holdsthe", "promise", "of", "contributing", "to", "more", "human", "-", "likecomputational", "models", "From", "the", "overall", "success", "ofpre", "-", "trained", "language", "models", "in", "metaphor", "detection", ",", "a", "certain", "degree", "of", "metaphoric", "knowledge", "in", "thesemodels", "can", "be", "assumed", "(", "Aghazadeh", "et", "al", ",2022", ")", ".This", "paper", "aims", "to", "evaluate", "whether", "this", "inherentknowledge", "extends", "beyond", "contextual", "clues", "to", "pre", "-", "dict", "the", "concrete", "domains", "in", "the", "metaphoric", "trans", "-", "fer", "Detecting", "a", "metaphor", "entails", "contrasting", "thephysical", "with", "the", "abstract", "meaning", "of", "a", "sequence", "However", ",", "the", "source", "domain", "is", "frequently", "a", "non", "-", "contextual", "attribute", "(", "Aghazadeh", "et", "al", ",2022", ")", ",", "whilethe", "target", "domain", "can", "be", "found", "directly", "using", "con", "-", "textual", "clues", "For", "instance", ",", "in", "the", "above", "example", ",", "pierceis", "more", "implicitly", "related", "to", "WEAPONSthan", "the", "explicitwordsis", "to", "WORDS", "To", "deter", "-", "mine", "the", "accuracy", "of", "the", "predicted", "source", "domainsfrom", "fine", "-", "tuning", "and", "few", "-", "shot", "prompting", "GPT-3", "(", "Brown", "et", "al", ",2020", ")", ",", "we", "manually", "evaluate", "the", "re", "-", "sults", "To", "this", "end", ",", "we", "propose", "a", "classification", "oferror", "types", "from", "too", "generic", "domains", "to", "relyingon", "words", "in", "the", "sentence", "that", "are", "not", "connected", "tothe", "metaphor", ",", "which", "we", "call", "trigger", "words", "Thisprovides", "further", "intuition", "on", "the", "nature", "and", "extentof", "metaphoric", "knowledge", "encoded", "in", "pre", "-", "trainedlanguage", "models", "We", "compare", "methods", "to", "elicitmetaphoric", "knowledge", "without", "any", "assumptionson", "grammar", "or", "source", "domains", "and", "test", "if", "it", "ex-1018", "tends", "across", "languages", ",", "i.e.", ",", "Spanish", "in", "addition", "toEnglish", "Finally", ",", "we", "evaluate", "its", "generalization", "bytesting", "on", "two", "distinct", "datasets", "and", "a", "set", "of", "non", "-", "metaphoric", "sentences.2", "PreliminariesTwo", "major", "pillars", "that", "build", "the", "foundation", "for", "thisapproach", "are", "conceptual", "metaphors", "and", "generativelanguage", "models", ",", "which", "we", "briefly", "introduce", "here.2.1", "Conceptual", "MetaphorsThe", "idea", "of", "metaphoric", "projection", "from", "a", "physi", "-", "cal", "source", "domain", "to", "an", "abstract", "target", "domain", "isdeeply", "rooted", "in", "the", "tradition", "of", "embodied", "cogni", "-", "tion", ",", "which", "assumes", "that", "higher", "-", "level", "cognition", "isshaped", "by", "physical", "experiences", "(", "Barsalou,1999", ")", ".For", "instance", ",", "actual", "physical", "movement", "recruitssimilar", "areas", "in", "the", "brain", "as", "communicating", "withaction", "verbs", "(", "Durand", "et", "al", ",2018", ";", "Gibbs,2006", ")", ".Conceptual", "metaphors", "are", "deeply", "entrenched", "inour", "knowledge", "organization", "system", "and", "utilized", "ineveryday", "communication", "to", "convey", "thoughts", "moreprecisely", "In", "a", "large", "-", "scale", "study", ",", "Prabhakaran", "et", "al.", "(", "2021", ")", "evaluate", "the", "persuasiveness", "of", "metaphorsand", "show", "that", "metaphoricity", "in", "political", "posts", "in", "-", "creases", "social", "media", "engagement", "Citron", "and", "Gold", "-", "berg", "(", "2014", ")", "show", "that", "metaphoric", "emotional", "lan", "-", "guage", "elicits", "a", "higher", "emotional", "response", "by", "recip", "-", "ients", "than", "literal", "use", "To", "provide", "complex", "analy", "-", "ses", "of", "metaphoricity", "in", "language", "and", "analyze", "themetaphoric", "knowledge", "of", "generative", "language", "mod", "-", "els", ",", "we", "believe", "that", "identifying", "concrete", "metaphoricprojections", "in", "natural", "language", "is", "required.2.2", "Generative", "Language", "ModelsLarge", "generative", "language", "models", "are", "trained", "withthe", "objective", "of", "predicting", "the", "next", "token", "in", "a", "se", "-", "quence", "During", "inference", ",", "this", "allows", "them", "to", "beprompted", "with", "some", "text", "by", "a", "user", "and", "then", "gen", "-", "erate", "what", "they", "predict", "to", "be", "most", "likely", "to", "comenext", "Scaled", "to", "large", "training", "corpora", "based", "onweb", "-", "data", "and", "multi", "-", "billion", "parameter", "architectures", ",", "this", "simple", "objective", "resulted", "in", "models", "such", "asGPT-3", "(", "Brown", "et", "al", ",2020", ")", "or", "its", "open", "-", "source", "vari", "-", "ants", "BLOOM", "(", "Luccioni", "et", "al", ",2022", ")", "and", "OPT-175", "(", "Zhang", "et", "al", ",2022", ")", "For", "a", "specific", "task", ",", "thesemodels", "can", "be", "used", "either", "in", "a", "zero", "-", "shot", ",", "few", "-", "shot", ",", "or", "fine", "-", "tuning", "manner", "For", "zero", "-", "shot", "text", "comple", "-", "tion", ",", "the", "model", "is", "prompted", "with", "an", "instance", "ofa", "task", "without", "being", "provided", "any", "example", "solu", "-", "tion", "of", "other", "task", "instances", "In", "comparison", ",", "forfew", "-", "shot", "completions", ",", "the", "prompt", "already", "containssome", "samples", "of", "the", "task", "and", "the", "respective", "solu", "-", "tions", "In", "both", "variants", ",", "the", "model", "weights", "are", "notchanged", "anymore", ",", "only", "the", "prompt", "differs", "In", "con", "-", "trast", ",", "when", "fine", "-", "tuning", "the", "model", ",", "its", "weights", "areoptimized", "to", "predict", "the", "task", "-", "specific", "output", "givensome", "input", "/", "output", "task", "samples.3", "Related", "WorkTong", "et", "al.", "(", "2021", ")", "provide", "a", "recent", "overview", "ofarchitectures", "used", "for", "metaphor", "detection", ",", "avail", "-", "able", "datasets", ",", "and", "further", "metaphor", "-", "related", "tasks", "An", "overview", "byRai", "and", "Chakraverty", "(", "2020", ")", "takes", "many", "different", "approaches", "to", "computationalmetaphor", "processing", "into", "account", ",", "additionally", "re-flecting", "on", "the", "different", "theoretical", "and", "linguisticviews", "on", "the", "definition", "of", "metaphors", "While", "thereare", "many", "metaphor", "-", "related", "tasks", ",", "the", "closest", "to", "oursare", "presented", "in", "the", "sections", "on", "paraphrasing", "andconnecting", "source", "and", "target", "domains", "Detection", "Metaphor", "detection", ",", "the", "simplest", "formof", "computational", "metaphor", "processing", ",", "is", "a", "binaryclassification", "task", "in", "which", "each", "word", "of", "a", "sentenceis", "labeled", "as", "being", "used", "metaphorically", "or", "literally", "In", "a", "2020", "shared", "task", "on", "metaphor", "detection", ",", "fine", "-", "tuning", "pre", "-", "trained", "language", "models", "led", "to", "the", "bestresults", "(", "Leong", "et", "al", ",2020", ")", "To", "achieve", "small", "im", "-", "provements", "in", "accuracy", ",", "different", "approaches", "en", "-", "rich", "the", "model", "input", "by", ",", "for", "instance", ",", "providingdictionary", "definitions", "of", "the", "words", "being", "classified", "(", "Babieno", "et", "al", ",2022", ")", "or", "concreteness", "measuresthat", "indicate", "to", "what", "extent", "something", "can", "be", "ex", "-", "perienced", "via", "the", "senses", "(", "Brysbaert", "et", "al", ",2014", ")", ".Commonly", "used", "datasets", "for", "this", "task", "are", "the", "VUAmsterdam", "Metaphor", "(", "VUA", ")", "Corpus", "(", "Steen", "et", "al", ",2010", ")", "and", "the", "TOEFL", "corpus", "(", "Klebanov", "et", "al", ",2018", ")", ",", "both", "human", "-", "annotated", "based", "on", "differentprotocols", "Model", "Insights", "Other", "research", "explores", "the", "em", "-", "beddings", "generated", "by", "language", "models", "and", "howthey", "relate", "to", "metaphoricity", "Pedinotti", "et", "al.", "(", "2021", ")", "show", "that", "BERT", "’s", "likelihood", "scores", "show", "a", "de", "-", "creasing", "likelihood", "from", "literal", "sentences", "to", "con", "-", "ventional", "and", "novel", "metaphors", "and", ",", "lastly", ",", "to", "non", "-", "sense", "sentences", ";", "thus", ",", "BERT", "’s", "scores", "correlate", "withhuman", "-", "annotated", "plausibility", "scores", "Moreover", ",", "for", "different", "layers", ",", "they", "explore", "MAE", "between", "words", "used", "metaphorically", ",", "e.g.", ",", "theflowersnoddedin", "the", "wind", ",", "and", "their", "metaphori", "-", "cal", "paraphrases", "and", "literal", "synonyms" ]
[ "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Similarly,1019", "Aghazadeh", "et", "al.", "(", "2022", ")", "investigate", "which", "layersof", "different", "language", "models", "encode", "metaphoricknowledge", "across", "different", "languages", "and", "datasetsvia", "probing", "Paraphrasing", "One", "common", "approach", "tometaphor", "interpretation", "is", "paraphrasing", "themetaphorical", "expression", "using", "only", "literal", "words", "For", "example", ",", "the", "phraseto", "devour", "a", "novelcouldbe", "rephrased", "asto", "enjoy", "a", "novel", "An", "exampleof", "metaphor", "interpretation", "is", "the", "work", "byMaoet", "al.", "(", "2018", ")", ",", "who", "propose", "to", "query", "WordNetfor", "possible", "candidate", "translations", ",", "from", "whichthe", "best", "is", "selected", "based", "on", "similarities", "in", "theembedding", "space", "On", "the", "other", "hand", ",", "there", "is", "alsoresearch", "on", "generating", "metaphoric", "paraphrasesgiven", "a", "literal", "sentence", "as", "input", "Recent", "work", "inmetaphoric", "paraphrasing", "uses", "text", "-", "to", "-", "text", "models", ",", "such", "as", "T5", "or", "BART", "(", "Stowe", "et", "al", ",2021b", ",", "a", ")", "Mostrecently", ",", "Liu", "et", "al.", "(", "2022", ")", "proposed", "a", "new", "task", "forwhich", "they", "created", "a", "dataset", "of", "novel", "metaphorsin", "the", "form", "of", "similes", ",", "for", "exampleThe", "meteorwas", "as", "bright", "as", "(", "New", "York", "City", "|", "coal", ")", ",", "whichthe", "language", "model", "has", "then", "to", "interpret", "asverybrightornot", "bright", "at", "all", "A", "fine", "-", "tuned", "RoBERTamodel", "outperforms", "various", "GPT", "variants", "on", "thetask", "and", "comes", "close", "to", "human", "performance", "Theauthors", "also", "show", "that", "the", "reverse", "of", "the", "tasks", ",", "i.e.", ",", "predicting", "the", "metaphoric", "language", "given", "theliteral", "answer", ",", "is", "more", "difficult", "Connecting", "Source", "and", "Target", "Domains", "Try", "-", "ing", "to", "automate", "the", "process", "of", "identifying", "metaphormappings", "is", "not", "a", "new", "endeavor", "For", "instance", ",", "given", "manually", "collected", "metaphoric", "phrases", "of", "aspecific", "target", "domain", ",", "Chung", "et", "al.", "(", "2004", ")", "proposeto", "facilitate", "the", "identification", "of", "source", "domains", "byquerying", "WordNet", "senses", "and", "the", "ontology", "SUMO.More", "recent", "research", "makes", "use", "of", "syntactical", "pat", "-", "terns", "metaphoric", "language", "often", "occurs", "in", "(", "Sullivan,2013", ")", ",", "thereby", "narrowing", "down", "the", "pool", "of", "sen", "-", "tences", "considered", "as", "metaphoric", "candidates", "Dodgeet", "al.", "(", "2015", ")", "use", "such", "patterns", "to", "find", "metaphorcandidates", "that", "are", "further", "analyzed", "by", "identifyingevoked", "frames", "and", "checking", "for", "whether", "the", "framesrelate", "in", "MetaNet", "Given", "a", "target", "domain", "and", "a", "cor", "-", "pus", ",", "they", "can", "use", "this", "system", "to", "see", "which", "sourcedomains", "are", "frequently", "used", "to", "metaphorically", "talkabout", "a", "target", "domain", "This", "system", ",", "however", ",", "islimited", "by", "existing", "frame", "resources", "and", "relies", "onpre", "-", "defined", "grammatical", "structures", "Also", "queryingan", "existing", "database", ",", "Ge", "et", "al.", "(", "2022", ")", "use", "hypernymrelations", "from", "WordNet", "to", "identify", "the", "source", "andtarget", "domains", "for", "pairs", "of", "literally", "used", "nouns", "andliterally", "or", "metaphorically", "used", "verbs", "or", "adjectives", "While", "the", "target", "domain", "identification", "reaches", "anaccuracy", "of", "87.3", "%", ",", "the", "source", "domain", "identifica", "-", "tion", "only", "reaches", "67.3", "%", "based", "on", "the", "manual", "eval", "-", "uation", "of", "a", "small", "subset", "of", "the", "data", "Shutova", "et", "al.", "(", "2017", ")", "explore", "unsupervised", "methods", "for", "identify", "-", "ing", "clusters", "of", "source", "and", "target", "concepts", "as", "wellas", "the", "connections", "between", "them", "They", "limit", "theirapproach", "to", "verb", "–", "noun", "constructions", ",", "from", "whichthe", "verbs", "constitute", "the", "source", "domain", "clusters", "andthe", "nouns", "the", "target", "domain", "clusters", "Mohler", "et", "al.", "(", "2016", ")", "provide", "a", "dataset", "withsentences", "from", "government", "discourse", "annotatedwith", "scores", "from", "-1", "to", "3", "to", "indicate", "the", "level", "ofmetaphoricity", "More", "importantly", ",", "7,941", "sentencesare", "annotated", "for", "source", "–", "target", "domain", "mappingswith", "108", "different", "source", "domains", "Rosen", "(", "2018", ")", "uses", "this", "dataset", "to", "build", "a", "model", "to", "predict", "thesource", "domain", "of", "a", "metaphor", "given", "a", "contextualsentence", "and", "a", "target", "domain", "referent", "Comparedto", "our", "approach", ",", "this", "work", "presupposes", "that", "a", "givensentence", "is", "metaphoric", "while", "also", "depending", "onspecific", "grammatical", "dependencies", "when", "construct", "-", "ing", "the", "model", "input", "Most", "importantly", ",", "it", "is", "limitedto", "the", "77", "labels", "sub", "-", "sampled", "from", "the", "overall", "avail", "-", "able", "108", "domains", "as", "experiments", "are", "done", "usingfeed", "-", "forward", "neural", "networks", "and", "LSTMs", "insteadof", "text", "-", "to", "-", "text", "networks", "Rosen", "also", "shows", "that", "theinter", "-", "annotator", "agreement", "for", "the", "original", "sourcedomain", "annotations", "is", "rather", "low", "with", "a", "Cohen’skappa", "of", "0.544", ",", "which", "indicates", "the", "difficulty", "andpotential", "ambiguity", "of", "the", "task", "In", "contrast", "to", "the", "existing", "work", "on", "computationalextraction", "of", "source", "and", "target", "domains", ",", "our", "ap", "-", "proach", "does", "not", "rely", "on", "any", "assumptions", "aboutgrammatical", "structure", "or", "word", "types", "that", "suppos", "-", "edly", "indicate", "metaphorical", "language", "Moreover", ",", "we", "are", "not", "limited", "to", "a", "pre", "-", "defined", "set", "of", "source", "ortarget", "domains", "due", "to", "the", "text", "-", "to", "-", "text", "approach.4", "Method4.1", "TaskIn", "our", "experiments", ",", "we", "use", "GPT-3", "to", "predict", "ametaphor", "’s", "source", "domain", "given", "a", "sentence", "anda", "target", "domain", "For", "example", ",", "a", "prompt", "to", "identifythe", "conceptual", "metaphor", "underlying", "the", "sentenceYou", "are", "wasting", "my", "timecould", "look", "like", "this", ":", "Extract", "the", "conceptual", "metaphor", "from", "thefollowing", "sentence", ":", "1020", "Sentence", ":", "Our", "relationship", "is", "atcrossroadsTarget", "Domain", ":", "RelationshipSource", "Domain", ":", "JourneySentence", ":", "You", "are", "wasting", "m", "y", "timeTarget", "Domain", ":", "TimeSource", "Domain", ":", "<", "<", "model", "completion", ">", ">", "In", "this", "prompt", ",", "the", "model", "is", "provided", "with", "oneexample", "of", "a", "metaphor", "mapping", ",", "which", "is", "RE", "-", "LATIONSHIP", "IS", "A", "JOURNEY", "Afterwards", ",", "it", "isprovided", "with", "the", "sentence", "and", "target", "domain", "forwhich", "we", "want", "to", "know", "the", "source", "domain", "A", "cor", "-", "rect", "prediction", ",", "in", "this", "case", ",", "would", "be", "TIME", "ISMONEY", "or", "TIME", "IS", "A", "RESOURCE.4.2", "DatasetThe", "main", "dataset", "was", "gathered", "by", "retrieving", "allnatural", "language", "examples", "annotated", "with", "sourceand", "target", "domain", "from", "Lakoff", "’s", "Master", "MetaphorList1", ",", "called", "Metaphor", "List", "in", "the", "following", "Forthis", "task", ",", "we", "randomly", "selected", "446", "sentences", ",", "witha", "maximum", "of", "three", "per", "metaphor", ",", "i.e.", ",", "per", "uniquecombination", "of", "source", "and", "target", "domain", "To", "en", "-", "sure", "that", "the", "model", "does", "not", "simply", "assume", "all", "sen", "-", "tences", "to", "be", "metaphoric", ",", "we", "use", "non", "-", "metaphoricEnglish", "sentences", "from", "the", "VUA", "corpus", "(", "Steenet", "al", ",2010", ")", "by", "extracting", "sentences", "for", "which", "eachword", "is", "labeled", "as", "literal", "by", "the", "annotators", "For", "in", "-", "stance", ",", "He", "did", "not", "even", "see", "an", "English", "newspaperis", "an", "example", "of", "a", "non", "-", "metaphoric", "sentence", "Fromthe", "extracted", "non", "-", "metaphoric", "sentences", ",", "we", "manu", "-", "ally", "chose", "50", "to", "be", "added", "to", "our", "dataset", "as", "many", "ofthe", "sentences", "were", "wrongly", "labeled", "or", "extremelyshort", "The", "resulting", "dataset", "is", "split", "into", "a", "train", ",", "vali", "-", "dation", ",", "and", "test", "set", "detailed", "in", "Table1", "A", "uniquecombination", "of", "source", "and", "target", "domain", ",", "for", "ex", "-", "ample", ",", "BELIEFS", "(", "target", ")", "ARE", "PLANTS", "(", "source", ")", ",", "does", "not", "appear", "in", "the", "validation", "or", "test", "set", "if", "italready", "appeared", "in", "the", "training", "set", "This", "allowsus", "to", "test", "whether", "the", "model", "can", "generalize", "to", "new", ",", "unseen", "metaphors", "As", "the", "Metaphor", "List", "data", "onlycontains", "a", "limited", "number", "of", "domain", "combinations", ",", "the", "validation", "and", "test", "set", "contain", "the", "same", "com", "-", "binations", "of", "source", "and", "target", "domains", ",", "however", ",", "with", "different", "unique", "sentences", "Entirely", "new", "do", "-", "main", "combinations", "in", "the", "test", "set", "are", "evaluated", "viasentences", "from", "additional", "datasets", "To", "test", "the", "ability", "to", "generalize", "across", "datasets,1http", ":", "/", "/", "www.lang.osaka", "-", "u.ac.jp", "/", "~sugimoto", "/", "MasterMetaphorList", "/", "metaphors", "/", "index.htmlcopyright", "(", "c", ")", "1994", "by", "George", "Lakoff", ",", "University", "ofCalifornia", ",", "Berkeleywe", "use", "sentences", "from", "the", "LCC", "dataset", "(", "Mohleret", "al", ",2016", ")", "(", "CC", "BY", "-", "NC", "-", "SA", "v4.0", ")", ",", "where", "we", "usethe", "provided", "source", "and", "target", "domains", "and", "theraw", "sentences", "without", "indication", "of", "the", "precisemetaphor", "location", "From", "the", "English", "and", "Span", "-", "ish", "sentences", ",", "we", "use", "a", "subset", "of", "maximally", "10sentences", "per", "target", "domain", ",", "resulting", "in", "a", "set", "of284", "(", "EN", ")", "and", "110", "(", "ES", ")", "sentences", "In", "comparisonto", "the", "Metaphor", "List", "samples", ",", "the", "LCC", "dataset", "con", "-", "sists", "of", "much", "longer", "sentences", "using", "complicated", ",", "expert", "language", "from", "the", "political", "domain", "All", "multilingual", "samples", ",", "as", "well", "as", "sentencesfrom", "the", "LCC", "corpus", ",", "are", "solely", "used", "as", "hold", "-", "outtest", "set", "and", "do", "not", "play", "a", "role", "in", "the", "model", "andprompt", "selection", "process", "These", "sentences", ",", "thus", ",", "test", "the", "model", "’s", "generalization", "ability", "to", "new", "sourcedomains", ",", "a", "different", "language", ",", "i.e.", ",", "Spanish", ",", "andmore", "complex", "sentences", "Model", "and", "prompt", "se", "-", "lection", "is", "based", "on", "the", "validation", "set", "created", "fromthe", "Metaphor", "List", "samples", "and", "the", "non", "-", "metaphoricVUA", "sentences", "The", "number", "of", "samples", "from", "thetraining", "set", "that", "are", "actually", "used", "depends", "on", "theprompting", "type.4.3", "Experiments", "and", "EvaluationUsing", "two", "automated", "evaluation", "metrics", "(", "Sec.4.3.1", ")", ",", "we", "compare", "few", "-", "shot", "prompts", "and", "fine", "-", "tuned", "models", "on", "the", "validation", "set", "(", "Sec.4.3.2", ")", "Thetest", "set", "evaluation", "is", "done", "manually", "(", "Sec.4.3.3", ")", "4.3.1", "Evaluation", "MetricWhile", "we", "manually", "evaluate", "the", "model", "on", "the", "testset", ",", "we", "use", "two", "automatically", "computed", "scores", "toevaluate", "on", "the", "validation", "set", "The", "validation", "per", "-", "formance", "is", "used", "to", "select", "the", "best", "way", "to", "promptor", "fine", "-", "tune", "the", "model", "As", "the", "first", "score", ",", "we", "com", "-", "pute", "the", "embedding", "similarity", "of", "the", "gold", "standardsource", "domain", "and", "the", "GPT-3", "-", "generated", "domain", "We", "compute", "the", "similarity", "using", "the", "Gensim", "library", "(", "ˇReh˚", "uˇrek", "and", "Sojka,2010", ")", "with", "300", "-", "dimensionalGloVe", "vectors", "(", "Pennington", "et", "al", ",2014", ")", ".To", "provide", "more", "context", "to", "the", "automated", "eval", "-", "uation", ",", "we", "also", "use", "knowledge", "graph", "embeddings", "We", "rely", "on", "the", "KGvec2go", "Web", "API", "(", "Portisch", "et", "al", ",2020", ")", "created", "from", "the", "resources", "WordNet", ",", "Wik", "-", "tionary", ",", "DBpedia", ",", "and", "WebIsALOD", "We", "averagethe", "four", "returned", "similarity", "scores", "based", "on", "the", "dif", "-", "ferent", "resources", ",", "called", "KB", "score", "in", "the", "following.4.3.2", "Prompt", "SelectionTo", "see", "with", "what", "prompts", "the", "model", "returns", "thebest", "source", "domains", ",", "we", "vary", "the", "number", "of", "labeled1021", "Dataset", "TrainSentencesVal", "Sen", "-", "tencesTest", "Sen", "-", "tencesTargetDomainsSourceDomainsMetaphor", "List", "117", "105", "224", "91", "94VUA", "non", "-", "metaphoric", "15", "15", "15", "47", "-LCC", "EN", "0", "0", "284", "30", "90LCC", "ES", "0", "0", "110", "11", "67Total", "132", "120", "633", "179", "251Table", "1", ":", "Number", "of", "sentences", "and", "unique", "target", "and", "source", "domains", "in", "the", "different", "datasets.few", "-", "shot", "samples", "provided", "at", "the", "beginning", "of", "eachprompt", "We", "compute", "the", "scores", "described", "in", "Sec", "-", "tion4.3.1for", "generations", "obtained", "through", "promptscontaining", "2", ",", "4", ",", "6", ",", "8", ",", "and", "12", "lr", "scheduler" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterValue", "O", "B-HyperparameterValue", "O", "B-HyperparameterValue", "O", "B-HyperparameterValue", "O", "O", "B-HyperparameterValue", "B-HyperparameterName", "I-HyperparameterName" ]
[ "Thatmeans", ",", "in", "each", "few", "-", "shot", "setting", ",", "the", "model", "has", "atleast", "2", "examples", "of", "correct", "domain", "mappings", "fororientation", "For", "each", "of", "these", "five", "prompt", "varia", "-", "tions", ",", "we", "choose", "three", "distinct", "sets", "of", "training", "sam", "-", "ples", "Thus", ",", "we", "generate", "three", "solutions", "to", "evaluate", "This", "allows", "us", "to", "observe", "how", "much", "the", "modeldepends", "on", "specific", "training", "samples", ",", "and", "we", "cancompute", "average", "scores", "and", "standard", "deviations", "Moreover", ",", "we", "also", "fine", "-", "tune", "GPT-3", "by", "using", "oursamples", "to", "train", "the", "model", "for", "4", "max", "seq", "length", ",", "duringwhich", "the", "model", "’s", "weights", "are", "adapted", ",", "instead", "ofjust", "providing", "the", "samples", "as", "optimizer", "inthe", "prompt" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterValue", "B-HyperparameterName", "I-HyperparameterName", "I-HyperparameterName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterName", "O", "O" ]
[ "After", "fine", "-", "tuning", ",", "the", "model", "does", "notrequire", "any", "epochs", "in", "the", "prompt", ",", "but", "candirectly", "classify", "a", "sample", "from", "the", "validation", "set" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "We", "fine", "-", "tune", "two", "variants", ":", "(", "1", ")", "a", "model", "fine", "-", "tunedwith", "all", "132", "sentences", "from", "the", "training", "set", ";", "and", "(", "2", ")", "a", "model", "fine", "-", "tuned", "with", "34", "sentences", "from", "thetraining", "set", ",", "one", "per", "unique", "source", "domain", "Theexperiments", "of", "comparing", "different", "prompts", "witheach", "other", "and", "the", "fine", "-", "tuned", "variants", "use", "the", "vali", "-", "dation", "set", "With", "GPT-3", "being", "proprietary", "licensedby", "the", "OpenAI", ",", "L.L.C", "Terms", "of", "Use", ",", "the", "text", "gen", "-", "erations", "with", "its", "API", "cost", "42.73", "$", "Our", "code", "isavailable", "online2.For", "all", "generations", ",", "we", "set", "the", "temperature", "pa", "-", "rameter", "to", "0", ",", "which", "means", "that", "the", "text", "generationmodel", "samples", "words", "in", "a", "greedy", "fashion", ",", "i.e.", ",", "it", "al", "-", "ways", "generates", "the", "most", "likely", "next", "word", "Increas", "-", "ing", "the", "temperature", "changes", "the", "likelihood", "withwhich", "words", "are", "sampled", "For", "now", ",", "a", "temperatureof", "0", "allows", "us", "to", "generate", "words", "in", "a", "deterministic", ",", "repeatable", "fashion", "However", ",", "future", "experimentscould", "include", "the", "temperature", "as", "a", "hyperparame", "-", "ter", "to", "be", "optimized", "The", "GPT-3", "architectures", "we2https", ":", "/", "/", "github.com", "/", "lwachowiak", "/", "Metaphor-Extraction-With-GPT-3used", "are", "davinci-002", ",", "the", "most", "powerful", "availablemodel", "variant", "at", "the", "time", "of", "the", "experiments3", ",", "andcurie-001", ",", "the", "second", "most", "powerful", "variant.4.3.3", "Manual", "EvaluationIssues", "with", "the", "gold", "standard", "source", "domains", ",", "aswell", "as", "the", "fact", "that", "the", "source", "domains", "can", "bephrased", "with", "different", "expressions", "and", "differ", "intheir", "level", "of", "precision", ",", "make", "it", "difficult", "for", "theautomated", "scores", "to", "be", "reliable", "enough", "to", "directlyderive", "an", "accuracy", "score", "from", "them", "Thus", ",", "to", "com", "-", "pute", "the", "final", "accuracy", "on", "the", "test", "set", ",", "we", "manu", "-", "ally", "check", "the", "model", "’s", "output", "After", "experimentingwith", "the", "different", "prompting", "styles", "on", "the", "validationset", ",", "we", "choose", "the", "model", "with", "the", "best", "Spearman", "and", "embedding", "similarity", "for", "manualevaluation", "on", "the", "hold", "-", "out", "test", "set" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterValue", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterValue", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Two", "annota", "-", "tors", ",", "the", "authors", "of", "this", "paper", ",", "manually", "evaluatethe", "correctness", "of", "the", "generated", "answers", "for", "En", "-", "glish", "Both", "annotators", "independently", "evaluated", "themodel", "output", "and", "then", "discussed", "disagreements", "One", "annotator", "evaluates", "the", "answers", "for", "Spanish", "The", "source", "domain", "was", "considered", "correct", "if", "itcorresponded", "to", "the", "gold", "standard", "or", "was", "deemedcorrect", "by", "the", "annotator", "(", "s", ")", ".While", "hard", "to", "automate", ",", "for", "humans", "it", "is", "ofteneasy", "to", "detect", "a", "close", "correspondence", "between", "agold", "domain", ",", "e.g", "“", "musical", "harmony", "”", ",", "and", "a", "pre", "-", "dicted", "domain", ",", "e.g", "“", "music", "”", "In", "difficult", "cases", ",", "annotators", ",", "following", "the", "Metaphor", "IdentificationProcedure", "(", "MIP", ")", "(", "Group,2007", ")", ",", "analyze", "words", "fortheir", "more", "basic", ",", "physical", "meaning", "and", "see", "if", "theseare", "in", "concordance", "with", "the", "predicted", "source", "do", "-", "main", "For", "instance", ",", "the", "gold", "standard", "forYou", "makeme", "sick", "!", "is", "“", "nausea", "”", ",", "whereassickis", "also", "definedas", "physically", "ill", "and", "thus", "related", "to", "the", "predicted“disease", "”", "domain", "To", "gather", "more", "insights", "intothe", "type", "of", "issues", "that", "can", "be", "observed", "from", "thepredicted", "source", "domains", ",", "all", "predictions", "deemed3In", "November", "2022", ",", "OpenAI", "released", "davinci-003", ",", "anInstructGPT", "variant", "(", "Ouyang", "et", "al", ",2022", ")", ".1022", "Figure", "1", ":", "Automatically", "computed", "scores", "on", "the", "valida", "-", "tion", "set", "in", "relation", "to", "the", "number", "of", "examples", "providedin", "the", "promptincorrect", "are", "classified", "by", "type", "of", "error", "as", "detailedin", "Section5.3.5", "ResultsThis", "section", "describes", "the", "results", "from", "the", "experi", "-", "ments", "that", "determine", "the", "manually", "evaluated", "testset", "predictions", ",", "their", "accuracy", ",", "and", "types", "of", "errors.5.1", "Prompt", "Selection", "ResultsFigure1shows", "the", "automatically", "computed", "scoreson", "the", "validation", "set", "achieved", "by", "davinci-002", "andcurie-001", "with", "different", "numbers", "of", "few", "-", "shot", "sam", "-", "ples", "We", "can", "see", "that", "davinci-002", "outperforms", "thesmaller", "architecture", "by", "about", "0.15", "to", "0.2", "points", "The", "highest", "embedding", "similarity", "and", "highest", "KBscore", "are", "achieved", "by", "davinci-002", "when", "promptedwith", "12", "different", "few", "-", "shot", "training", "samples", ",", "achiev", "-", "ing", "an", "embedding", "similarity", "of", "0.505", "and", "a", "KBscore", "of", "0.553", "However", ",", "the", "standard", "deviation", "isvery", "high", "for", "the", "models", "prompted", "with", "12", "samples", ",", "thus", ",", "showing", "the", "importance", "of", "the", "quality", "of", "thosesamples", "Due", "to", "this", "fluctuation", "in", "performance", ",", "the", "average", "KB", "score", "over", "all", "three", "runs", "is", "highestfor", "davinci-002", "models", "prompted", "with", "8", "samples", ",", "and", "the", "average", "embedding", "score", "is", "highest", "fordavinci-002", "models", "prompted", "with", "4", "samples", "Theprompt", "based", "on", "12", "batch", "size", "that", "led", "tothe", "overall", "best", "results", "is", "available", "in", "the", "appendix" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterValue", "B-HyperparameterName", "I-HyperparameterName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "In", "comparison", ",", "the", "fine", "-", "tuned", "models", "performbetter", "than", "the", "curie-001", "models", "but", "worse", "thanthe", "davinci-002", "models", "Fine", "-", "tuning", "a", "model", "with36", "samples", ",", "each", "with", "a", "unique", "source", "domain", ",", "leads", "to", "an", "MSE", "of", "0.303", "and", "MCC", "of", "0.386" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "B-HyperparameterValue", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "B-MetricValue", "O", "B-MetricName", "O", "B-MetricValue" ]
[ "Fine", "-", "tuning", "GPT-3", "on", "all", "avail", "-", "able", "training", "samples", "results", "in", "improved", "scores", "of0.413", "and", "0.513", "Examining", "the", "completions", "of", "themodel", "fine", "-", "tuned", "on", "all", "samples", ",", "we", "can", "see", "that", "itsticks", "more", "to", "the", "source", "domains", "already", "presentin", "the", "training", "data", "while", "also", "predicting", "fewerdistinct", "source", "domains", "overall", ":", "the", "completionsfrom", "the", "best", "performing", "few", "-", "shot", "variant", "contain74", "unique", "source", "domains", ",", "from", "which", "7", "are", "alsopresent", "in", "the", "training", "data", ";", "the", "completions", "of", "themodel", "fine", "-", "tuned", "on", "36", "contain", "78", "unique", "sourcedomains", ",", "from", "which", "13", "are", "present", "in", "the", "trainingdata", ";", "and", "the", "completions", "of", "the", "model", "fine", "-", "tunedon", "all", "data", "contain", "only", "50", "unique", "source", "domains", ",", "from", "which", "18", "are", "present", "in", "the", "training", "data.5.2", "Manual", "Evaluation", "ResultsWe", "used", "the", "best", "prompt", "identified", "in", "the", "previoussection", "to", "generate", "the", "source", "domains", "for", "the", "testset", "samples", "The", "correctness", "of", "the", "generationswas", "manually", "verified", "by", "two", "annotators", "We", "BLEU", ",", "a", "chance", "-", "corrected", "coefficient", "ofagreement", ",", "to", "compute", "the", "inter", "-", "annotator", "agree", "-", "ment" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Across", "all", "test", "data", "points", ",", "we", "obtained", "a", "Co", "-", "hen", "’s", "Kappa", "of", "0.51", ",", "corresponding", "to", "a", "moderateagreement", "according", "toLandis", "and", "Koch", "(", "1977", ")", ".After", "disagreements", "were", "resolved", "through", "discus", "-", "sion", ",", "we", "computed", "the", "model", "’s", "accuracy", ",", "which", "isreported", "by", "dataset", "in", "Table2", "The", "model", "achievedan", "accuracy", "of", "81.33", "%", "on", "the", "Metaphor", "List", "cor", "-", "pus", ",", "53.74", "%", "on", "the", "English", "part", "of", "the", "LCC", "corpus", ",", "and", "34.65", "%", "on", "the", "Spanish", "part", "of", "the", "LCC", "cor", "-", "pus", "In", "addition", ",", "the", "model", "was", "able", "to", "achieve", "anaccuracy", "of", "42.11", "%", "in", "predicting", "a", "sentence", "is", "non", "-", "metaphoric", "instead", "of", "predicting", "a", "source", "domain", "Averaged", "by", "sample", ",", "this", "results", "in", "an", "accuracy", "of60.22", "%", "The", "decrease", "in", "performance", "on", "the", "LCCtest", "set", "is", "not", "surprising", "as", "the", "sentences", "are", "on", "av", "-", "erage", "much", "longer", "and", "often", "use", "domain", "-", "specificlanguage", "Moreover", ",", "the", "target", "domains", "specifiedby", "the", "LCC", "gold", "standard", "are", "often", "much", "harder", "toidentify", "in", "the", "sentence", "as", "they", "are", "less", "preciselymatched", "to", "the", "sentence", "’s", "words", "To", "provide", "insights", "into", "the", "adequacy", "of", "the", "eval", "-", "uation", "metrics", ",", "we", "evaluate", "their", "correlation", "withthe", "manual", "annotation", "decisions", "As", "we", "have", "anordinal", "variable", "(", "correctly", "classified", ",", "wrongly", "clas", "-", "sified", ")", "and", "a", "continuous", "variable", "(", "KB", "score", "andembedding", "similarity", ")", ",", "we", "used", "Spearman", "’s", "rankcorrelation", "coefficient", "We", "achieve", "a", "correlation", "of0.43", "for", "the", "KB", "score", "and", "0.40", "for", "the", "embeddingsimilarity", "Trigger", "here", "refers", "to", "words", "inthe", "input", "that", "are", "clearly", "related", "to", "the", "predictedsource", "domain", "For", "instance", ",", "any", "mentions", "ofanimal", "-", "related", "terms", ",", "e.g.bullish", "mindsetortroughof", "poverty", ",", "led", "the", "model", "to", "predict", "“", "animals", "”", "assource", "domain", "The", "most", "common", "error", "class", "is", "be", "-", "ing", "wrong", "without", "any", "trigger", "in", "the", "sentence", ",", "fol", "-", "lowed", "by", "erroneous", "predictions", "of", "non", "-", "metaphoricand", "being", "wrong", "with", "trigger", "Some", "instances", "in", "-", "dicate", "a", "misinterpretation", "of", "words", ",", "e.g.dumb", "-", "foundedlikely", "leads", "to", "the", "entertaining", "predictionof", "“", "being_stupid", "”", "Furthermore", ",", "interesting", "errorscan", "be", "found", "in", "the", "category", "of", "wrong", "subelementmappings", ",", "where", "the", "model", "identifies", "the", "generalsource", "domain", "but", "fails", "to", "pick", "the", "correct", "elementof", "that", "domain", "for", "its", "prediction", "For", "instance", ",", "in", "the", "sentenceChina", "is", "a", "fertile", "ground", "for", "re", "-", "volt", ",", "the", "gold", "standard", "refers", "to", "“", "plants", "”", ",", "and", "themodel", "predicts", "“", "land", "”", ",", "which", "is", "in", "the", "same", "domainof", "cultivation", "but", "not", "entirely", "the", "correct", "domain", "Similarly", ",", "when", "a", "metaphor", "involved", "movementand", "locations", "and", "the", "true", "source", "domain", "referredto", "only", "one", "of", "them", ",", "the", "model", "regularly", "pickedthe", "wrong", "subelement", "For", "instance", ",", "the", "modelwrongly", "predicts", "EXISTANCE", "IS", "MOTION", "forthe", "sentenceIt", "came", "into", "existence", ",", "where", "the", "truesource", "domain", "would", "have", "been", "“", "location”", "For", "the", "Spanish", "LCC", "data", ",", "one", "annotator", "classi-fied", "erroneous", "predictions", "according", "to", "our", "errortypology", "A", "vast", "majority", "of", "62.12", "%", "of", "errors", "werepredictions", "of", "non", "-", "metaphoric", "sequences", "whichshould", "be", "metaphoric", ",", "followed", "by", "19.70", "%", "wrongwithout", "trigger", "A", "trend", "to", "predict", "“", "family", "”", "withoutany", "trigger", "in", "the", "sentence", "for", "the", "target", "domain“government", "”", "in", "half", "of", "its", "occurrences", "could", "beobserved", "In", "the", "13.64", "%", "cases", "of", "wrong", "with", "trig", "-", "ger", ",", "the", "model", "’s", "predictions", "mostly", "represented", "lit", "-", "eral", "English", "translations", "of", "context", "words", "from", "theSpanish", "sentence", "All", "source", "domain", "predictionswere", "made", "in", "English", ",", "which", "was", "expected", "giventhat", "the", "source", "and", "target", "domains", "in", "the", "promptwere", "also", "in", "English", "In", "total", ",", "12", "LCC", "sentenceswere", "disregarded", "since", "the", "gold", "standard", "was", "faulty.6", "DiscussionWe", "experimented", "with", "different", "GPT-3", "variants", "andprompts", "containing", "varying", "numbers", "of", "few", "-", "shotsamples", "to", "see", "whether", "GPT-3", "can", "generate", "thesource", "domain", "of", "a", "conceptual", "metaphor", "mappinggiven", "a", "context", "and", "a", "target", "domain", "The", "best", "re", "-", "sults", "were", "achieved", "with", "a", "long", "few", "-", "shot", "promptcontaining", "12", "example", "completions", "The", "largestmodel", "variant", "davinci-002", "strongly", "outperformedthe", "next", "biggest", "variant", "and", "a", "fine", "-", "tuned", "GPT-3.We", "also", "saw", "that", "fine", "-", "tuning", "the", "model", "canlead", "to", "a", "decrease", "in", "expressiveness", ",", "that", "is", ",", "fewerunique", "source", "domains", "being", "generated", "In", "ourcase", ",", "this", "might", "be", "because", "the", "model", "fine", "-", "tunedon", "all", "data", "sees", "each", "source", "domain", "around", "threetimes", "per", "training", "It", "might", "be", "possible", "to", "counter", "-", "act", "the", "decrease", "in", "expressiveness", "by", "increasing", "thetemperature", "parameter", ",", "thus", ",", "making", "less", "probablegenerations", "more", "likely", "Manually", "coding", "the", "errors", "made", "by", "the", "model", ",", "we", "saw", "that", "the", "model", "often", "fabricates", "source", "do", "-", "mains", "for", "which", "no", "related", "words", "are", "present", "in", "thesentence", "Other", "common", "errors", "included", "predictinga", "literal", "meaning", "although", "a", "metaphor", "was", "present", ",", "and", "generating", "wrong", "source", "domains", "based", "ontrigger", "words", "that", "were", "not", "metaphorically", "related1024", "Error", "Code", "Definition", "Example", "%", "of", "AllErrorsSentence", "Wrong", "PredictionWrong", "with", "trig", "-", "gerThe", "model", "predicts", "awrong", "source", "domain", "dueto", "words", "in", "the", "sentence", "re", "-", "lated", "to", "that", "domainThe", "arms", "raceCOMPETITION", "ISWAR21.31Wrong", "withouttriggerThe", "model", "predicts", "awrong", "source", "domainwithout", "any", "noticeabletriggers", "for", "that", "domain", "inthe", "sentenceSally", "gave", "the", "ideato", "SamIDEAS", "ARE", "CHIL", "-", "DREN27.32Too", "literalThe", "model", "predicts", "a", "lit", "-", "eral", "relationship", "instead", "ofa", "metaphoric", "mappingI’m", "down", "to", "my", "bot", "-", "tom", "dollarMONEY", "IS", "IN", "-", "VESTMENT7.10Should", "be", "non", "-", "metaphoricThe", "model", "predicted", "ametaphoric", "source", "domaininstead", "of", "non", "-", "metaphoricThey", "saw", "him", "ad", "-", "vancingMOVING", "IS", "COM", "-", "ING7.65Should", "bemetaphoricThe", "model", "wrongly", "pre", "-", "dicted", "non", "-", "metaphoricUnder", "the", "cover", "ofdarknessDARKNESS", "ISnon", "-", "metaphoric25.14Too", "specificThe", "predicted", "metaphor", "ismore", "specific", "than", "whatthe", "sentence", "implies", "He", "finally", "caughtup", "to", "scheduleSCHEDULE", "ISPEOPLE2.73Too", "generalThe", "predicted", "source", "do", "-", "main", "is", "too", "unspecificThe", "idea", "slippedthrough", "my", "fingersMIND", "IS", "SPACE", "1.09Wrong", "subele", "-", "ment", "mappingThe", "model", "predicts", "an", "as", "-", "pect", "of", "the", "correct", "sourcedomain", ",", "however", ",", "it", "is", "notthe", "exact", "elementLet", "’s", "strip", "away", "theunimportant", "detailsIMPORTANCE", "ISCLOTHING7.65Table", "3", ":", "The", "different", "types", "of", "errors", "made", "by", "the", "modelto", "the", "target", "domain", "Discerning", "whether", "to", "predicta", "source", "domain", "for", "a", "given", "sentence", "or", "to", "label", "itas", "non", "-", "metaphoric", "seems", "to", "be", "quite", "challengingfor", "the", "model", "as", "well", "Analyzing", "the", "errors", "of", "largelanguage", "models", "as", "done", "here", "is", "essential", "to", "buildappropriate", "trust", "or", "distrust", "in", "the", "model", "and", "allowfor", "the", "use", "of", "error", "-", "correction", "methods", "in", "the", "fu", "-", "ture", ",", "for", "instance", ",", "the", "selection", "of", "better", "prompts", "ortraining", "samples", "In", "the", "context", "of", "analyzing", "the", "model", "’s", "misclas", "-", "sifications", ",", "we", "also", "experienced", "issues", "with", "thedataset", ",", "e.g", "unintuitive", "metaphor", "mappings", "orlack", "of", "contextual", "clues", "for", "the", "provided", "targetdomain", "The", "dataset", "’s", "quality", "strongly", "affectedthe", "Spanish", "test", "results", "and", "clearly", "indicated", "thatmore", "multilingual", "resources", "for", "metaphor", "identi-fication", "are", "needed", "The", "difference", "in", "the", "natureand", "quality", "of", "the", "datasets", "is", "also", "the", "main", "reasonfor", "the", "strong", "variation", "in", "ROUGE" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName" ]
[ "TheMetaphor", "List", "dataset", "provides", "prototypical", ",", "gen", "-", "eral", "language", "examples", ",", "while", "the", "LCC", "datasetannotated", "real", "-", "world", ",", "domain", "-", "specific", "expert", "lan", "-", "guage", "This", "affects", "the", "complexity", "as", "well", "as", "thelength", "of", "sentences", ",", "both", "contributing", "to", "the", "differ", "-", "ence", "in", "accuracy", "across", "datasets", "Application", "Using", "GPT-3", "to", "analyze", "metaphorsused", "in", "an", "unlabeled", "corpus", "comes", "with", "two", "prob", "-", "lems", ":", "(", "1", ")", "we", "do", "not", "know", "what", "target", "domains", "arethe", "right", "ones", "to", "provide", "to", "the", "model", ",", "(", "2", ")", "there", "willbe", "an", "overwhelming", "amount", "of", "output", "given", "that1025", "most", "sentences", "contain", "at", "least", "subtle", "metaphoriclanguage", "that", "will", "largely", "not", "even", "be", "relevant", "to", "thedomain", "we", "are", "interested", "in", "Therefore", ",", "it", "would", "beuseful", "to", "first", "filter", "sentences", "based", "on", "seed", "wordswhose", "usage", "interests", "us", "or", "that", "belong", "to", "a", "spe", "-", "cific", "target", "domain", "we", "want", "to", "analyze", "(", "Wachowiaket", "al", ",2022", ")", "As", "such", "an", "approach", "already", "narrowsdown", "the", "candidate", "sentences", "to", "a", "pre", "-", "specifiedtarget", "domain", ",", "we", "can", "include", "that", "target", "domainin", "the", "prompt", "for", "the", "language", "model", "Lastly", ",", "itmight", "help", "to", "restrict", "the", "context", "window", "aroundthe", "words", "of", "interest", "so", "that", "the", "model", "is", "not", "dis", "-", "tracted", "by", "other", "metaphors", "in", "the", "sentence", "How", "-", "ever", ",", "to", "confirm", "this", ",", "further", "research", "is", "needed", "Considering", "precise", "element", "mappings", "As", "thecapabilities", "of", "large", "neural", "language", "models", "con", "-", "tinue", "to", "grow", ",", "it", "will", "be", "interesting", "to", "see", "if", "they", "canidentify", "not", "only", "the", "correct", "source", "domains", "butalso", "precise", "element", "-", "wise", "mappings", "between", "theconcepts", "of", "the", "target", "and", "source", "domain", "For", "ex", "-", "ample", ",", "the", "conceptual", "metaphor", "LOVE", "IS", "A", "JOUR", "-", "NEY", "involves", "mapping", "lovers", "to", "travelers", ",", "difficul", "-", "ties", "to", "roadblocks", ",", "and", "progress", "to", "distance", "traveledforward", "Querying", "such", "an", "element", "-", "wise", "mappingcould", "be", "facilitated", "through", "a", "set", "of", "the", "target", "do", "-", "main", "’s", "core", "elements", "being", "provided", "to", "the", "model", "OpenAI", "Transparency", "Issues", "An", "issue", "with", "re", "-", "searching", "the", "capabilities", "of", "large", "language", "models", ",", "such", "as", "GPT-3", ",", "is", "the", "accessibility", "and", "transparency", "While", "GPT-3", "variants", "are", "easily", "accessible", "via", "anAPI", ",", "the", "model", "stays", "a", "black", "-", "box", ",", "and", "researcherscan", "not", "investigate", "the", "specific", "model", "weights", "Moreover", ",", "there", "is", "no", "explicit", "mapping", "available", "ofhow", "the", "models", "advertised", "on", "the", "website", "relate", "tothose", "described", "in", "OpenAI", "’s", "papers", "(", "Leike,2022", ")", ".Lastly", ",", "the", "model", "variant", "accessible", "for", "fine", "-", "tuningdiffers", "from", "the", "one", "accessible", "for", "direct", "zero-", "andfew", "-", "shot", "text", "generation", ",", "which", "might", "also", "explainthe", "drop", "in", "performance", "observed", "in", "our", "metaphorextraction", "task", "On", "the", "other", "hand", ",", "comparablemodels", "for", "which", "the", "weights", "are", "publicly", "released", ",", "such", "as", "BLOOM", "(", "Luccioni", "et", "al", ",2022", ")", "or", "OPT-175", "(", "Zhang", "et", "al", ",2022", ")", ",", "have", "the", "issue", "that", "theyare", "not", "hosted", "anywhere", "Thus", ",", "researchers", "mustprovide", "the", "infrastructure", "to", "run", "them", ",", "which", "isonly", "possible", "for", "very", "few", "academic", "institutions.7", "ConclusionWe", "analyzed", "how", "well", "GPT-3", "can", "identify", "sourcedomains", "of", "metaphors", "in", "natural", "language", "Acrossthree", "different", "datasets", "in", "English", "and", "Spanish", ",", "GPT-3", "predicts", "the", "source", "domain", "with", "an", "accuracyof", "60.22", "%", "The", "best", "performance", "was", "achievedgiven", "12", "epochs", "in", "the", "prompt", ",", "al", "-", "though", "the", "average", "performance", "was", "highest", "with4", "to", "8", "attention", "type" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterValue", "B-HyperparameterName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterValue", "O", "B-HyperparameterValue", "B-HyperparameterName", "I-HyperparameterName" ]
[ "However", ",", "the", "modelstill", "suffers", "from", "specific", "error", "types", ",", "such", "as", "hal", "-", "lucinating", "domains", "without", "any", "indicators", "beingpresent", "We", "believe", "future", "iterations", "of", "large", "lan", "-", "guage", "models", "like", "GPT-3", "will", "become", "importanttools", "in", "computational", "metaphor", "analysis", ",", "whereone", "investigates", "conceptual", "metaphors", "in", "differentdomains", ",", "for", "instance", ",", "literature", "or", "political", "dis", "-", "course", "In", "the", "future", ",", "we", "want", "to", "experiment", "withusing", "large", "language", "models", "to", "generate", "completemetaphors", ",", "i.e.", ",", "generate", "both", ",", "source", "domain", "andtarget", "domain", ",", "given", "a", "sentence", "We", "also", "plan", "touse", "the", "developed", "techniques", "in", "corpus", "analyses", "LimitationsThe", "approach", "of", "identifying", "source", "domains", "relieson", "having", "a", "contextual", "sentence", "but", "also", "a", "targetdomain", "available", "The", "datasets", "available", "for", "evalu", "-", "ation", "do", "not", "always", "provide", "precise", "target", "domains", "For", "example", ",", "the", "LCC", "dataset", "provides", "the", "targetdomaingun", "ownershipfor", "the", "sentenceI", "just", "don’tknow", "what", "it", "will", "take", "for", "people", "in", "this", "country", "toembrace", "gun", "safety", ",", "or", "the", "target", "domainclimatechangefor", "the", "sentenceThe", "event", "is", "billed", "as", "thelargest", "meeting", "of", "influential", "figures", "within", "the", "re", "-", "newable", "energy", "field", "This", "mismatch", "often", "makesit", "difficult", "to", "provide", "precise", "source", "domains", "Asimilar", "problem", "also", "exists", "when", "wanting", "to", "useour", "source", "domain", "prediction", "approach", "in", "the", "wildas", "we", "have", "to", "somehow", "provide", "the", "model", "witha", "target", "domain", "While", "we", "can", "provide", "a", "targetdomain", "by", "selecting", "sentences", "based", "on", "seed", "-", "wordlists", "designed", "for", "specific", "domains", ",", "we", "do", "not", "knowhow", "precisely", "this", "matches", "the", "target", "domain", "occur", "-", "ring", "in", "the", "sentence" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]