sentences
sequence
ner_tags
sequence
[ "Proceedings", "of", "the", "61st", "Annual", "Meeting", "of", "the", "Association", "for", "Computational", "Linguistics", "Volume", "1", ":", "Long", "Papers", ",", "pages", "1–21", "July", "9", "-", "14", ",", "2023", "©", "2023", "Association", "for", "Computational", "Linguistics", "One", "Can", "not", "Stand", "for", "Everyone", "!", "Leveraging", "Multiple", "User", "Simulators", "to", "train", "Task", "-", "oriented", "Dialogue", "Systems", "Yajiao", "LIU1,2", ",", "Xin", "Jiang3", ",", "Yichun", "Yin3", ",", "Yasheng", "Wang3", ",", "Fei", "Mi3", "Qun", "Liu3", ",", "Xiang", "Wan2", ",", "Benyou", "Wang1,2", "1The", "Chinese", "University", "of", "Hong", "Kong", ",", "Shenzhen", "2Shenzhen", "Research", "Institute", "of", "Big", "Data", "3Huawei", "Noah", "’s", "Ark", "Lab", "yajiaoliu@link.cuhk.edu.cn", "Abstract", "User", "simulators", "are", "agents", "designed", "to", "imi-", "tate", "human", "users", ";", "recent", "advances", "have", "found", "that", "RNN", ")", ",", "see", "Fig", ".", "1(a", ")", ".", "The", "reason", "might", "be", "that", "these", "user", "simulators", "(", "i.e.", ",", "with", "either", "handcrafted", "rules", "or", "data", "-", "driven", "learning", "in", "their", "DM", "modules", ")", "have", "different", "user", "dialog", "act", "distributions4(see", "Fig", ".", "1(b", ")", ")", "which", "determine", "the", "dialogue", "state", "space", "explored", "by", "the", "dialogue", "system", ".", "One", "can", "not", "stand", "for", "everyone", ".", "Users", "might", "behave", "differently", ",", "one", "could", "design", "different", "user", "simulators", "with", "specific", "user", "dialog", "act", "distributions", ",", "see", "Shi", "et", "al", ".", "(", "2019", ")", ".", "A", "single", "user", "simulator", "learned", "on", "a", "task", "-", "oriented", "dialogue", "corpus", "can", "just", "represent", "one", "or", "a", "group", "of", "users", ",", "while", "the", "dialogue", "system", "needs", "to", "accomplish", "tasks", "from", "various", "human", "users", "in", "real", "scenarios", ".", "We", "argue", "that", "it", "is", "beneficial", "to", "utilize", "all", "different", "user", "simu-", "the", "system", "agent", "tested", "with", "user", "simulator", "Uj", ",", "and", "Tj", ",", "tis", "the", "num-", "ber", "of", "times", "user", "simulator", "Ujhas", "been", "selected", "with", "so", "far", "." ]
[ "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", "B-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterName", "O", "O", "O", "B-TaskName", "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", "B-MethodName", "I-MethodName", "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", "B-MethodName", "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", "B-MethodName", "O", "O", "B-MethodName", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O" ]
[ "Then", "we", "normalize", "}", ")", ")", ",", "(", "3", ")", "Eq", ".", "3", "penalizes", "the", "user", "simulators", "with", "which", "the", "dialogue", "system", "already", "performs", "well", "in", "the", "ex-", "pectation", "term", "." ]
[ "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" ]
[ "Where", "the", "hyperparameter", "τis", "the", "smooth", "factor", "for", "distribution", "p={p1,···,pK", "}", "–", "the", "larger", "τis", ",", "the", "sharper", "pis", "." ]
[ "O", "O", "O", "B-HyperparameterName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterName", "O", "O", "O", "B-HyperparameterName", "O" ]
[ "Each", "probability", "pjinpis", "calculated", "as", "pj", "=", "zj", "/", "summationtextK", "i=1zi", ".", "(", "4", ")" ]
[ "O", "O", "B-HyperparameterName", "O", "O", "B-HyperparameterName", "O", "B-HyperparameterValue", "O", "B-HyperparameterValue", "B-HyperparameterValue", "O", "O", "O", "O" ]
[ "In", "the", "following", "T−1dialogues", ",", "we", "will", "spec-", "ify", "all", "user", "simulators", "to", "train", "the", "system", "agent", "S", "with", "this", "distribution", "p", "(lines", "15", "-", "18", ")", "." ]
[ "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "B-HyperparameterName", "O", "O", "O", "O", "O", "O" ]
[ "We", ";", "Compan-", "ion", "Volume", ",", "Short", "Papers", ",", "pages", "149–152", ",", "Rochester", ",", "New", "York", ".", "Association", "for", "Computational", "Linguis-", "tics", ".", "Jost", "Schatzmann", "and", "Steve", "Young", ".", "2009", ".", "The", "hidden", "agenda", "user", "simulation", "model", ".", "IEEE", "Transactions", "on", "Audio", ",", "Speech", ",", "and", "Language", "Processing", ",", "17(4):733", "–", "747", ".", "Weiyan", "Shi", ",", "Kun", "Qian", ",", "Xuewei", "Wang", ",", "and", "Zhou", "Yu", ".", "2019", ".", "How", "to", "build", "user", "simulators", "to", "train", "RL", "-", "based", "dialog", "systems", ".", "In", "Proceedings", "of", "the", "2019", "Confer-", "ence", "on", "Empirical", "Methods", "in", "Natural", "Language", "Processing", "and", "the", "9th", "International", "Joint", "Conference", "on", "Natural", "Language", "Processing", "(", "EMNLP", "-", "IJCNLP", ")", ",", "pages", "1990–2000", ",", ",", "China", "of", "turns", "of", "the", "dialogue", "exceeds", "10", ".", "The", "reward", "will", "be", "given", "+1", "for", "task", "success", ",", "-1", "for", "task", "failure", ",", "and", "-0.1", "for", "each", "additional", "turn", "to", "encourage", "the", "RL", "-", "based", "policy", "module", "to", "finish", "the", "task", "fast", ".", "Also", ",", "a", "discounted", "factor", "of", "0.9", "is", "applied", "to", "all", "the", "experiences", ".", "B.4", "The", "parameters", "of", "training", "Sys", "-", "MUST", "adaptive", "The", "hyperparameters", "used", "to", "train", "the", "Sys", "-", "MUST", "adaptive", "are", "listed", "in", "the", "Tab", ".", "5", ".", "Since", "some", "user", "simulators", "used", "for", "implement-", "ing", "our", "MUST", "framework", "are", "based", "on", "the", "GPT", "model", ",", "we", "train", "Sys", "-", "MUST", "adaptive", "on", "a", "V100", "GPU", "and", "it", "will", "cost", "around", "15", "hours", "with", "the", "default", "hyperparameters", "above", ".", "B.5", "Human", "Evaluation", "on", "dialogue", "systems", "We", "find", "5", "volunteers", "to", "conduct", "the", "human", "evaluations", "on", "dialogue", "systems", ".", "They", "all", "have", "good", "English", "skills", "and", "the", "exploration", "term", "in", "our", "modified", "UCB1", "algorithm", ",", "the", "simplest", "way", "to", "calculate", "the", "distribution", "pis", "to", "make", "the", "sample", "probability", "w.r.t", "a", "user", "simulator", "solely", "depend", "on", "the", "inversion", "of", "the", "system", "’s", "performance", ".", "See", "the", "row", "called", "‘", "w", "/", "t", "exploration", "’", "in", "Tab", ".", "10", "for", "comparisons", "." ]
[ "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-DatasetName", "I-DatasetName", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "I-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "B-DatasetName", "I-DatasetName", "O", "B-DatasetName", "I-DatasetName", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "B-TaskName", "B-TaskName", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "B-TaskName", "B-TaskName", "O", "B-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", "B-MetricValue", "O", "O", "O", "O", "O", "O", "O", "O", "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-HyperparameterName", "O", "O", "O", "O", "O", "B-HyperparameterName", "O", "O", "B-HyperparameterName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterValue", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "B-MethodName", "B-MethodName", "O", "O", "B-HyperparameterName", "O", "O", "B-MethodName", "O", "O", "O", "B-HyperparameterValue", "B-HyperparameterName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterName", "O", "O", "B-MethodName", "O", "O", "O", "B-MethodName", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "B-MethodName", "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" ]
[ "In", "this", "situation", ",", "the", "obtained", "distribution", "p", "might", "be", "sharp", "due", "to", "the", "lack", "of", "the", "exploration", "term", ",", "which", "would", "be", "harmful", "for", "uniform", "adaption", "to", "some", "extent", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "As", "Fig", ".", "5(a", ")", "shows", ",", "MUST", "adaptive", "performs", "worse", "and", "con-", "verges", "slower", "when", "omitting", "the", "exploration", "term", ",", "compared", "with", "when", "our", "modified", "UCB1", "algorithm", "has", "the", "exploration", "term", ".", "This", "could", "demonstrate", "both", "the", "importance", "of", "uniform", "adaption", "and", "the", "usefulness", "of", "the", "exploration", "term", "." ]
[ "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", "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" ]
[ "E.2", "Ablation", "study", "on", "the", "designed", "distribution", "a", "post", "-", "processing", "for", "the", "re", "-", "scaling", "purpose", "and", "a", "sum", "-", "one", "normalization", ".", "Since", "we", "omit", "the", "exploration", "term", "for", "the", "second", "row", ",", "therefore", ",", "it", "does", "not", "need", "the", "post", "-", "processing", ".", "MUST", "adaptive", "-", "III", "moves", "the", "re", "-", "scaling", "and", "the", "inversion", "operations", "to", "the", "front", "of", "the", "constructed", "exploitation", "-", "exploration", "term", ".", "our", "design", "with", "three", "operations", "(", "i.e.", ".", ",", "exploitation", "-", "exploration", "term", "construction", ",", "re", "-", "scaling", "strategy", ",", "and", "the", "inversion", "of", "¯xj", ")", "is", "not", "only", "reasonable", "but", "also", "robust", "to", "the", "order", "permutation", "of", "these", "three", "operations", "." ]
[ "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "B-TaskName", "O", "B-TaskName", "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", "B-MethodName", "B-MethodName", "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" ]
[ "F", "Implementing", "MUST", "with", "more", "user", "simulators", "To", "implement", "our", "MUST", "with", "more", "user", "simula-", "tors", ",", "we", "use", "Simulated", "Agenda", "Dataset", "to", "train", "four", "extra", "user", "simulators8", ".", "Fig", ".", "6(a", ")", "shows", "the", "learn-", "ing", "curve", "of", "the", "system", "agent", "trained", "by", "GPT", "GPT", "RNN", "RNN", "GPT-RNN", "GPT-RNN", "RL", "RL", "U-GPT", "U-GPT", "NLU", "NLU", "NLG", "NLG", "POL", "POL", "RNNT", "RNNT", "RNNR", "RNNR", "GPT-RNN", "GPT-RNN", "GPT", "GPT", "Sys-RNNT", "Sys-RNNT", "Sys-RNNR", "Sys-RNNR", "Sys-RNN", "Sys-RNN", "Sys-GPT-RNN", "Sys-GPT-RNN", "RNNT", "RNNT", "RNNR", "RNNR", "RNN", "RNN", "GPT-RNN", "GPT-RNN", "RNNT", "RNNT", "RNNR", "RNNR", "RNN", "RNN", "GPT-RNN", "GPT-RNN", "ACL", "2023", "MNLI", "Multi", "GLUE", "CoNLL-2003", "access", "conditions", "(", "in", "particular", ",", "derivatives", "of", "data", "accessed", "for", "research", "purposes", "should", "not", "be", "used", "outside", "of", "research", "contexts", ")", "?", "No", "response", ".", "/squareB4", ".", "Did", "you", "discuss", "the", "steps", "taken", "to", "check", "whether", "the", "data", "that", "was", "collected", "/", "used", "contains", "any", "information", "that", "names", "or", "uniquely", "identifies", "individual", "people", "or", "offensive", "content", ",", "and", "the", "steps", "taken", "to", "protect", "/", "anonymize", "it", "?", "No", "response", ".", "/squareB5", ".", "Did", "you", "provide", "documentation", "of", "the", "artifacts", ",", "e.g.", ",", "coverage", "of", "domains", ",", "languages", ",", "and", "linguistic", "phenomena", ",", "demographic", "groups", "represented", ",", "etc", ".", "?", "No", "response", ".", "/squareB6", ".", "Did", "you", "report", "relevant", "statistics", "like", "the", "number", "of", "examples", ",", "details", "of", "train", "/", "test", "/", "dev", "splits", ",", "etc", ".", "for", "the", "data", "that", "you", "used", "/", "created", "?", "Even", "for", "commonly", "-", "used", "benchmark", "datasets", ",", "include", "the", "number", "of", "examples", "in", "train", "/", "validation", "/", "test" ]
[ "B-MethodName", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-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", "B-MethodName", "I-MethodName", "B-MethodName", "I-MethodName", "B-MethodName", "I-MethodName", "B-MethodName", "I-MethodName", "B-MethodName", "I-MethodName", "B-MethodName", "I-MethodName", "B-MethodName", "I-MethodName", "B-MethodName", "I-MethodName", "B-MethodName", "I-MethodName", "B-MethodName", "I-MethodName", "B-MethodName", "I-MethodName", "B-MethodName", "I-MethodName", "B-MethodName", "I-MethodName", "B-MethodName", "I-MethodName", "B-MethodName", "I-MethodName", "B-MethodName", "I-MethodName", "B-MethodName", "I-MethodName", "B-MethodName", "I-MethodName", "B-MethodName", "I-MethodName", "B-MethodName", "I-MethodName", "B-MethodName", "I-MethodName", "B-MethodName", "I-MethodName", "B-MethodName", "I-MethodName", "B-MethodName", "I-MethodName", "B-DatasetName", "I-DatasetName", "B-DatasetName", "B-DatasetName", "B-DatasetName", "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", "B-HyperparameterName", "O", "B-HyperparameterName", "O", "O", "O", "O", "O", "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", "O", "B-HyperparameterName", "O", "B-HyperparameterName" ]
[ "Proceedings", "of", "the", "61st", "Annual", "Meeting", "of", "the", "Association", "for", "Computational", "Linguistics", "Volume", "1", ":", "Long", "Papers", ",", "pages", "22–35", "July", "9", "-", "14", ",", "2023", "©", "2023", "Association", "for", "Computational", "Linguistics", "SAFECONV", ":", "Explaining", "and", "Correcting", "Conversational", "Unsafe", "Behavior", "Lab", ",", "Bellevue", ",", "WA", ",", "USA", "{", "Abstract", "One", "of", "the", "main", "challenges", "open", "-", "domain", "end", "-", "to", "-", "end", "dialogue", "systems", ",", "or", "chatbots", ",", "face", "is", "the", "prevalence", "of", "unsafe", "behavior", ",", "such", "as", "toxic", "languages", "and", "harmful", "suggestions", ".", "However", ",", "existing", "dialogue", "datasets", "do", "not", "provide", "enough", "prepares", "canned", "responses", "as", "safe", "alternatives", ".", "However", ",", "the", "canned", "responses", "are", "just", "one", "of", "two", "types", "of", "safe", "contextual", "-", "irrelevant", "utterances", ".", "We", "propose", "contextual", "rewriting", ",", "a", "new", "way", "to", "generate", "safe", ",", "diverse", ",", "and", "context", "-", "relevant", "alternative", "responses", "given", "the", "context", "and", "unsafe", "response", ".", "As", "shown", "in", "Figure", "1", ",", "the", "alternative", "response", "produced", "by", "contextual", "rewriting", "is", "a", "better", "choice", "to", "replace", "the", "unsafe", "response", ",", "improving", "coherence", "and", "contextual", "relevance", "of", "the", "response", ".", "However", ",", "no", "datasets", "provide", "explicit", "supervision", "on", "how", "to", "respond", "nicely", "and", "toxicity", "-", "free", "while", "conforming", "to", "the", "conversational", "context", "when", "unsafe", "behavior", "occurs", ".", "To", "tackle", "the", "above", "issues", ",", "we", "propose", "SAFECONV", ",", "a", "large", "-", "scale", "dataset", "of", "dialogues", "for", "the", "research", "of", "conversational", "safety", ",", "where", "(", "1", ")", "which", "removes", "dialogues", "containing", "toxic", "words", "and", "sensitive", "content", ",", "and", "classifier", "-", "based", "filtering", ",", "which", "filters", "out", "dialogues", "regarding", "sensitive", "topics", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "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", "B-TaskName", "B-TaskName", "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", "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" ]
[ "PchatbotW", "sourced", "their", "dialogues", "crawled", "from", "Weibo", ",", "however", ",", "compared", "to", "LCCC", ",", "their", "data", "cleaning", "procedures", "relating", "to", "toxicity", "are", "not", "as", "comprehensive", ":", "they", "only", "filter", "dialogues", "with", "sensitive", "words", ".", "Therefore", ",", "PchatbotW", "contains", "more", "frequent", ",", "explicit", "unsafe", "behavior", "while", "for", "LCCC", "-", "base", ",", "more", "infrequent", "and", "implicit", ",", "which", "we", "call", "the", "safety", "-", "graduated", "attribute", "of", "SAFECONV", "." ]
[ "B-MethodName", "O", "O", "O", "O", "O", "B-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", "B-MethodName", "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-MethodName", "O" ]
[ "Moreover", ",", "the", "dialogues", "from", "two", "sources", "differ", "in", "content", "types", ",", "with", "LCCC", "-", "base", "containing", "mainly", "daily", "conversation", "and", "PchatbotW", "havingmore", "cases", "of", "comments", "over", "a", "post", ",", "such", "as", "a", "news", "headline", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "We", "verify", "the", "safety", "-", "graduated", "attribute", "by", "a", "trained", "safety", "checker", "(", "see", "subsection", "3.2", "are", "safe", "on", "the", "surface", "but", "express", "toxicity", "or", "cause", "serious", "risks", "with", "reference", "to", "the", "context", ",", "such", "as", "agreement", "to", "suicide", "or", "harmful", "medical", "advice", ";", "they", "are", "usually", "a", "whole", "sentence", "or", "a", "clause", ",", "rather", "than", "just", "a", "toxic", "word", ",", "such", "as", "Why", "not", "?", "(", "#", "6", ")", "in", "Table", "2", ".", "Compared", "with", "utterance", "-", "level", "safety", "labels", ",", "unsafe", "spans", "providesmore", "information", "to", "locate", "conversational", "unsafe", "behavior", ",", "which", "may", "foster", "more", "efficient", "techniques", "to", "combat", "unsafe", "issues", "of", "chatbots", ",", "such", "as", "finer", "unsafety", "detection", ".", "Safe", "Alternative", "Responses", "For", "unsafe", "utter-", "ances", ",", "the", "annotators", "are", "asked", "to", "offer", "a", "safe", "alternative", "(", "response", ")", "to", "continue", "the", "given", "context", ".", "The", "safe", "alternatives", "are", "supposed", "to", "correct", "the", "occurred", "unsafe", "behavior", "and", "guide", "the", "conversa-", "tion", "to", "move", "towards", "a", "safe", "and", "linear", "binary", "classification", "encoder", "CLS", "SEP", "CLS", "SEP", "BART", ")", "1988", "283", "(", "%", "85.8", "⇓", ")", "67", "(", "%", "96.7", "⇓", ")", "Table", "6", ":", "Results", "of", "explainable", "checking", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "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", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "B-TaskName", "B-TaskName", "B-TaskName", "B-TaskName", "B-HyperparameterName", "B-HyperparameterName", "B-HyperparameterName", "B-HyperparameterName", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "6", "Correct", "Conversational", "Unsafe", "Behavior", "via", "Contextual", "Rewriting", "One", "solution", "to", "avoid", "unsafe", "behavior", "is", "to", "conduct", "a", "check", "-", "reject", "-", "regenerate", "cycle", "—", "checking", "the", "gen-", "erated", "response", "with", "a", "safety", "checker", ",", "refusing", "it", "if", "it", "is", "unsafe", ",", "and", "regenerating", "a", "new", "response", "—", "repeatedly", "until", "a", "safe", "response", "surfaces", ".", "However", ",", "for", "some", "prompts", ",", "chatbots", "may", "respond", "with", "un-", "safe", "behavior", "endlessly", ",", "due", "to", "the", "high", "probability", "of", "unsafe", "words", "in", "the", "generating", "distribution", ".", "A", "more", "efficient", "method", "is", "one", "-", "time", "checking", "and", "rewriting", "—", "directly", "rewriting", "unsafe", "responses", "into", "detoxified", "ones", "with", "a", "rewriter", "trained", "on", "unsafesafe", "response", "pairs", ".", "However", ",", "no", "dataset", "could", "support", "a", "satisfactory", "rewriter", "in", "the", "past", ".", "Correspondingly", ",", "the", ")", "\"", ".", "Flue", ".", "Cohe", ".", "Info", ".", "Unsafe", "Before", "Rewriting", "3.27", "2.27", "2.85", "92.6", "%", "After", "Rewriting", "3.25", "2.29", "2.75", "36.5", "%", "After", "Finetuning", "3.38", "2.39", "2.79", "9.7", "%", "Table", "8", ":", "Human", "evaluation", "of", "the", "responses", ".", "question", ":", "can", "we", "further", "improve", "the", "rewriter", "by", "making", "it", "aware", "of", "its", "bad", "generations", "?", "We", "further", "finetune", "the", "rewriter", "on", "the", "feedback", "of", "the", "safety", "checker", "with", "PPO", "(", "Schulman", "et", "al", ".", ",", "2017", ";", "Ouyang", "et", "al", ".", ",", "2022", ")", ",", "a", "policy", "optimization", "method", "in", "Reinforcement", "Learning", "(", "RL", ")", ".", "Specifically", ",", "the", "objective", "to", "optimize", "is", ":", "J(θ", ")", "=", "E(x", ",", "y′)∼R", "θ[r(x", ",", "y′)−βlogRθ(y′|x", ")", "Rθ′(y′|x", ")", "]", ",", "where", "θandθ′are", "the", "parameters", "of", "the", "rewriter", "to", "optimize", "and", "before", "finetuning", ";", "x", ",", "yandy′denote", "our", "dataset", "to", "collect", "unsafe", "prompts", ",", "responses", ",", "or", "spans", "and", "mis-", "use", "them", ".", "This", "is", "a", "common", "issue", "for", "all", "public", "datasets", "regarding", "toxicity", "or", "safety", ".", "We", "believe", "that", "our", "dataset", "creates", "more", "value", "than", "risks", ".", "Besides", ",", "there", "is", "no", "leakage", "of", "personal", "information", "because", "our", "data", "sources", ",", "LCCC", "-", "base", "(", "Wang", "et", "al", ".", ",", "2020", ")", "and", "PchatbotW", "(", "Qian", "et", "al", ".", ",", "2021", ")", "have", "al-", "ready", "been", "preprocessed", "to", "remove", "personal", "in-", "formation", "by", "researchers", "of", "previous", "work", "(", "see", "their", "papers", "for", "details", ")", ".", "Also", ",", "though", "our", "dataset", "contains", "more", "instances", "compared", "to", "previously", "proposed", "datasets", ",", "the", "dialogues", "are", "mostly", "from", "social", "media", "and", "may", "not", "cover", "types", "of", "conversational", "unsafe", "behavior", "found", "in", "other", "media", ".", "All", "the", "procedure", "and", "rules", "to", "collect", "SAFECONV", "are", "approved", "by", "the", "ethics", "review", "committee", "at", "Linguistics", ".", "Samuel", "Gehman", ",", "Suchin", "Gururangan", ",", "Maarten", "Sap", ",", "Yejin", "Choi", ",", "and", "Noah", "A.", "Smith", ".", "2020", ".", "RealToxi-", "cityPrompts", ":", "Evaluating", "neural", "toxic", "degeneration", "in", "language", "models", ".", "In", "Findings", "of", "the", "Association", "for", "Computational", "Linguistics", ":", "EMNLP", "2020", ",", "pages", "3356–3369", ",", "Online", ".", "Association", "for", "Computational", "Linguistics", ".", "Amelia", "Glaese", ",", "Nat", "McAleese", ",", ",", "et", "al", ".", "1", "(", ")", ",", "pages", "1702–1723", ",", ".", ".", "2021", ".", "“", "nice", "try", ",", "kiddo", "”", ":", "Investigating", "ad", "hominems", "in", "dialogue", "responses", ".", "In", ",", "pages", "750–767", ",", "that", "you", "used", "/", "created", "?", "Even", "for", "commonly", "-", "used", "benchmark", "datasets", ",", "include", "the", "number", "of", "examples", "in", "train", "/", "validation", "/", "test", "splits", ",", "as", "these", "provide", "necessary", "context", "for", "a", "reader", "to", "understand", "experimental", "results", ".", "For", "example", ",", "small", "differences", "in", "accuracy", "on", "large", "test", "sets", "may", "be", "significant", ",", "while", "on", "small", "test", "sets", "they", "may", "not", "be", ".", "Left", "blank", ".", "C", "/", "square", "Did", "you", "run", "computational", "experiments", "?", "4,5,6", "/", "square", "C1", ".", "Did", "you", "report", "the", "number", "of", "parameters", "in", "the", "models", "used", ",", "the", "total", "computational", "budget", "(", "e.g.", ",", ",", "GPU", "hours", ")", ",", "and", "computing", "infrastructure", "used", "?", "4,5,6", "The", "Responsible", "NLP", "Checklist", "used", "at", "ACL", "2023", "is", "adopted", "from", "NAACL", "2022", ",", "with", "the", "addition", "of", "a", "question", "on", "AI", "writing", "assistance.34", "/", "square", "C2", ".", "Did", "you", "discuss", "the", "experimental", "setup", ",", "including", "hyperparameter", "search", "and" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "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", "B-HyperparameterName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "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", "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", "B-HyperparameterName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "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", "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", "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", "B-TaskName", "B-TaskName", "B-TaskName", "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", "B-HyperparameterValue", "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", "B-HyperparameterName", "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", "B-MethodName", "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", ":", ",", "pages", "36–50", ",", "2023", "©", "2023", "Association", "for", "Computational", "Linguistics", "Detecting", "and", "Mitigating", "Hallucinations", "in", "Machine", "Translation", ":", "Model", "Internal", "Workings", "Alone", "Do", "Well", ",", "Sentence", "Similarity", "Even", "Better", "{", "daviddale,lenavoita,loicbarrault,costajussa}@meta.com", "Abstract", "While", "the", "problem", "of", "hallucinations", "in", "neural", "machine", "translation", "has", "long", "been", "recognized", ",", "so", "far", "the", "progress", "on", "its", "alleviation", "is", "very", "little", ".", "Indeed", ",", "recently", "it", "turned", "out", "that", "with-", "out", "artificially", "encouraging", "models", "to", "halluci-", "nate", ",", "previously", "existing", "methods", "fall", "short", "and", "even", "hallucinations", "with", "precision", "80", "%", "higher" ]
[ "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", "B-TaskName", "I-TaskName", "B-TaskName", "O", "B-TaskName", "B-TaskName", "O", "B-TaskName", "I-TaskName", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "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", "B-MetricName", "B-MetricName", "B-MetricName", "B-MetricValue", "B-MetricValue", "O" ]
[ "ALTI", "AUC", "84.9", "Hallucination", "detection", "COMET-QE", "AUC", "70.2", "Hallucination", "detection", "LASER", "AUC", "79.4", "Hallucination", "detection", "LaBSE", "AUC", "91.7", "Hallucination", "detection", "XNLI", "AUC", "90.9", "Hallucination", "detection", ":", "quality", ".", "Metrics", ":", "ROC", "AUC", "(", "↑", ")", "and", "P@R90", "(", "↑", ")", ".", "Methods", ":", "oracle", ",", "internal", ",", "external", ".", "Changes", "in", "scores", "are", "highlighted", "compared", "to", "Seq", "-", "Logprob", ".", "both", "all", "and", "fully", "detached", "hallucinations", ",", "their", "pre", "-", "cision", "at", "90", "%", "recall", "is", "roughly", "twice", "better", "than", "that", "of", "Seq", "-", "Logprob", ".", "While", "such", "a", "good", "performance", "might", "be", "expected", "for", "LaBSE", "that", "evaluates", "cross", "-", "lingual", "sentence", "similarity", "(", "in", "a", "way", ",", "this", "might", "be", "seen", "as", "a", "measure", "of", "translation", "quality", ")", ",", "results", "for", "XNLI", "are", "rather", "surprising", ":", "to", "the", "best", "of", "our", "knowledge", ",", "models", "optimized", "for", "XNLI", "have", "not", "been", "used", "in", "the", "context", "of", "machine", "translation", ".", "Note", "also", "the", "large", "difference", "between", "LaBSE", "and", "LASER", ":", "while", "the", "former", "shows", "big", "improve", "-", "ments", "compared", "to", "Seq", "-", "beam", "search", "size", "5", "beam", "search", "size", "N", "80", "beam", "widths", "1,", "3,", "10" ]
[ "B-MethodName", "B-MetricName", "B-MetricValue", "B-TaskName", "B-TaskName", "B-MethodName", "B-MetricName", "B-MetricValue", "B-TaskName", "B-TaskName", "B-MethodName", "B-MetricName", "B-MetricValue", "B-TaskName", "B-TaskName", "B-MethodName", "B-MetricName", "B-MetricValue", "B-TaskName", "B-TaskName", "B-MethodName", "B-MetricName", "B-MetricValue", "B-TaskName", "B-TaskName", "O", "O", "O", "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", "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", "B-MetricName", "B-MetricName", "B-MetricName", "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-TaskName", "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", "B-MethodName", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterName", "B-HyperparameterName", "B-HyperparameterName", "B-HyperparameterValue", "B-HyperparameterName", "B-HyperparameterName", "B-HyperparameterName", "B-HyperparameterValue", "B-HyperparameterValue", "B-HyperparameterName", "B-HyperparameterName", "B-HyperparameterValue", "B-HyperparameterValue", "B-HyperparameterValue" ]
[ "diversity", "rates", "1,", "3,", "10" ]
[ "B-HyperparameterName", "B-HyperparameterName", "B-HyperparameterValue", "B-HyperparameterValue", "B-HyperparameterValue" ]
[ "Monte", "Carlo", "dropout", "iterations", "5", "COMET" ]
[ "B-HyperparameterName", "B-HyperparameterName", "B-HyperparameterName", "B-HyperparameterName", "B-HyperparameterValue", "B-MetricName" ]
[ "BLEU", "LaBSE", "XNLI", "hallucination", "detection" ]
[ "B-MetricName", "B-MetricName", "B-MetricName", "B-TaskName", "B-TaskName" ]
[ "translation", "quality", "evaluation" ]
[ "B-TaskName", "B-TaskName", "B-TaskName" ]
[ "manual", "annotation" ]
[ "B-TaskName", "B-TaskName" ]
[ "automatic", "evaluation" ]
[ "B-TaskName", "B-TaskName" ]
[ "BEAM", "10", "COMET", "10", "COMET", "LaBSE", "ALTI", "LASER", "LaBSE", "XNLI", "ALTI", "LASER", "LaBSE", "XNLI", "COMET", "COMET", "COMET", "COMET", "COMET", "ALTI", "LASER", "LaBSE", "XNLI", "LaBSE", "ALTI", "LaBSE", "ALTI", "LaBSE", "ALTI", "LaBSE", "ALTI", "LaBSE", "ALTI", "LaBSE", "ALTI", "LaBSE", "ALTI", "LaBSE", "ALTI", "LaBSE", "ALTI", "LaBSE", "ALTI", "LaBSE", "ALTI", "LaBSE", "ALTI", "COMET", ":", "Evaluating", "cross-", "lingual", "sentence", "representations", ".", "In", "Proceedings", "of", "the", "2018", "Conference", "on", "Empirical", "Methods", "in", "Natural", "Language", "Processing", ".", "Association", "for", "Computational", "Linguistics", ".", "Fangxiaoyu", "Feng", ",", "Yinfei", "Yang", ",", "Daniel", "Cer", ",", "Language", "-", "agnostic", "BERT", "sentence", "embedding", ".", "In", "Proceedings", "of", "the", "60th", "Annual", "Meeting", "of", "the", "Association", "for", "Computational", "Linguistics", "(", "Volume", "1", ":", "Long", "Papers", ")", ",", "pages", "878–891", ",", "Dublin", ",", "Ireland", ".", "Association", "for", "Computa-", "tional", "Linguistics", ".", "Patrick", ",", ":", "A", "framework", "for", "salient", "long", "-", "tail", "translation", "error", "detection", ".", "2020a", ".", "COMET", ":", "A", "neural", "framework", "for", "MT", "evaluation", ".", "In", "Proceedings", "of", "the", "2020", "Conference", "on", "Empirical", "Methods", "in", "Natural", "Language", "Processing", "(", "EMNLP", ")", ",", "pages", "2685–2702", ",", "Online", ".", "Association", "for", "Computational", "Linguistics", ".", ",", "and", "’s", "participation", "in", "the", "WMT20", "metrics", "shared", "task", ".", "In", "Proceedings", "of", "the", "Fifth", "Conference", "on", "Machine", "Translation", ",", "pages", "911–920", ",", "On-", "line", ".", "Association", "for", "Computational", "Linguistics", ".", "Katsuhito", "A", "A1", "A2", "A3", "A4", "B", "B1", "B2", "B3", "B4", "B5", "B6", "C", "C1", "C2", "C3", "C4", "NLTK", "Spacy", "ROUGE", "Did", "you", "use", "human", "annotators", "(", "e.g.", ",", "crowdworkers", ")", "or", "research", "with", "human", "participants", "?", "section", "5", "/square", "D1", ".", "Did", "you", "report", "the", "full", "text", "of", "instructions", "given", "to", "participants", ",", "including", "e.g.", ",", "screenshots", ",", "disclaimers", "of", "any", "risks", "to", "participants", "or", "annotators", ",", "etc", ".", "?", "appendix", "C", "/square", "D2", ".", "Did", "you", "report", "information", "about", "how", "you", "recruited", "(", "e.g.", ",", "crowdsourcing", "platform", ",", "students", ")", "and", "paid", "participants", ",", "and", "discuss", "if", "such", "payment", "is", "adequate", "given", "the", "participants", "’", "demographic", "(", "e.g.", ",", "country", "of", "residence", ")", "?", "The", "annotators", "were", "members", "of", "our", "team", "and", "did", "the", "job", "within", "their", "normal", "working", "hours", ".", "/square", "D3", ".", "Did", "you", "discuss", "whether", "and", "how", "consent", "was", "obtained", "from", "people", "whose", "data", "you", "’re", "using", "/", "curating", "?", "For", "example", ",", "if", "you", "collected", "data", "via", "crowdsourcing", ",", "did", "your", "instructions", "to", "crowdworkers" ]
[ "B-MethodName", "B-HyperparameterValue", "B-MetricName", "B-HyperparameterValue", "B-MetricName", "B-MethodName", "B-MethodName", "B-MethodName", "B-MethodName", "B-MethodName", "B-MethodName", "B-MethodName", "B-MethodName", "B-MethodName", "B-MetricName", "B-MetricName", "B-MetricName", "B-MetricName", "B-MetricName", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "B-MetricValue", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "B-TaskName", "B-TaskName", "O", "O", "O", "B-TaskName", "B-TaskName", "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", "B-TaskName", "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", "B-MethodName", "O", "O", "O", "O", "O", "B-TaskName", "B-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "B-TaskName", "B-TaskName", "O", "B-MethodName", "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", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "B-MethodName", "I-HyperparameterName", "I-HyperparameterName", "I-HyperparameterName", "I-HyperparameterName", "B-MethodName", "I-HyperparameterName", "I-HyperparameterName", "I-HyperparameterName", "I-HyperparameterName", "I-HyperparameterName", "I-HyperparameterName", "B-MethodName", "I-HyperparameterName", "I-HyperparameterName", "I-HyperparameterName", "I-HyperparameterName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "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", "78–91", "July", "9", "-", "14", ",", "2023", "©", "2023", "Association", "for", "Computational", "Linguistics", "Span", "-", "Selective", "Linear", "Attention", "Transformers", "for", "Effective", "and", "Robust", "{", "Abstract", "In", "estimate", "the", "current", "state", "of", "a", "conversation", "using", "natural", "language", "descriptions", "of", "the", "ser-", "vice", "schema", "for", "generalization", "to", "unseen", "ser-", "vices", ".", "Prior", "generative", "approaches", "which", "decode", "slot", "values", "sequentially", "do", "not", "generalize", "well", "to", "variations", "in", "schema", "," ]
[ "O", "O", "O", "O", "O", "O", "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", "B-TaskName", "O", "O", "B-TaskName", "B-TaskName", "B-TaskName", "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" ]
[ "SPLAT", "N", "value:", "Metric:", "Intent", "Accuracy", "Metric", "value:", "2.5", "Task:", "Named", "Entity", "Recognition", "Schema-Guided", "Dialogue", "(SGD)" ]
[ "I-MethodName", "I-HyperparameterName", "O", "B-MetricName", "I-MetricName", "I-MetricName", "B-MetricName", "B-MetricValue", "I-MetricValue", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-DatasetName", "I-DatasetName", "I-DatasetName" ]
[ "Method:", "SPLAT", "Hyperparameter:", "L", "Hyperparameter", "value:", "Metric:", "Joint", "Goal", "Accuracy", "(JGA)", "Metric", "value:", "7", "Task:", "Named", "Entity", "Recognition", "Dataset:", "Schema-Guided", "Dialogue", "(SGD)", "Method:", "SPLAT", "Hyperparameter:", "M", "Hyperparameter", "value:", "-", "Metric:", "Intent", "Accuracy", "Metric", "value:", "30", "Task:", "Named", "Entity", "Recognition", "Dataset:", "Schema-Guided", "Dialogue", "(SGD)" ]
[ "B-MethodName", "I-MethodName", "B-HyperparameterName", "I-HyperparameterName", "B-HyperparameterName", "B-HyperparameterValue", "B-MetricName", "I-MetricName", "I-MetricName", "I-MetricName", "I-MetricName", "B-MetricName", "B-MetricValue", "I-MetricValue", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "B-DatasetName", "I-DatasetName", "I-DatasetName", "I-DatasetName", "B-MethodName", "I-MethodName", "B-HyperparameterName", "I-HyperparameterName", "B-HyperparameterName", "O", "O", "B-MetricName", "I-MetricName", "I-MetricName", "B-MetricName", "B-MetricValue", "I-MetricValue", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "B-DatasetName", "I-DatasetName", "I-DatasetName", "I-DatasetName" ]
[ "Method:", "SPLAT", "Hyperparameter:", "Mj", "Hyperparameter", "value:", "-", "Metric:", "Joint", "Goal", "Accuracy", "(JGA)", "Metric", "value:", "6", "Task:", "Named", "Entity", "Recognition", "Dataset:", "Schema-Guided", "Dialogue", "(SGD)" ]
[ "B-MethodName", "I-MethodName", "B-HyperparameterName", "I-HyperparameterName", "B-HyperparameterName", "O", "O", "B-MetricName", "I-MetricName", "I-MetricName", "I-MetricName", "I-MetricName", "B-MetricName", "B-MetricValue", "I-MetricValue", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "B-DatasetName", "I-DatasetName", "I-DatasetName", "I-DatasetName" ]
[ "Method:", "SPLAT", "Hyperparameter:", "IletR", "Hyperparameter", "value:", "-", "Metric:", "Intent", "Accuracy", "Metric", "value:", "2021", "Task:", "Named", "Entity", "Recognition", "Dataset:", "Schema-Guided", "Dialogue", "(SGD)" ]
[ "B-MethodName", "I-MethodName", "B-HyperparameterName", "I-HyperparameterName", "B-HyperparameterName", "O", "O", "B-MetricName", "I-MetricName", "I-MetricName", "B-MetricName", "B-MetricValue", "I-MetricValue", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "B-DatasetName", "I-DatasetName", "I-DatasetName", "I-DatasetName" ]
[ "Method:", "SPLAT", "Hyperparameter:", "M", "Hyperparameter", "value:", "-", "Metric:", "Joint", "Goal", "Accuracy", "(JGA)", "Metric", "value:", "2021", "Task:", "Named", "Entity", "Recognition", "Dataset:", "Schema-Guided", "Dialogue", "(SGD)" ]
[ "B-MethodName", "I-MethodName", "B-HyperparameterName", "I-HyperparameterName", "B-HyperparameterName", "O", "O", "B-MetricName", "I-MetricName", "I-MetricName", "I-MetricName", "I-MetricName", "B-MetricName", "B-MetricValue", "I-MetricValue", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "B-DatasetName", "I-DatasetName", "I-DatasetName", "I-DatasetName" ]
[ "Method:", "SPLAT", "Hyperparameter:", "R1", "Hyperparameter", "value:", "-", "Metric:", "Intent", "Accuracy", "Metric", "value:", "30", "Task:", "Named", "Entity", "Recognition", "Dataset:", "Schema-Guided", "Dialogue", "(SGD)" ]
[ "B-MethodName", "I-MethodName", "B-HyperparameterName", "I-HyperparameterName", "B-HyperparameterName", "O", "O", "B-MetricName", "I-MetricName", "I-MetricName", "B-MetricName", "B-MetricValue", "I-MetricValue", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "B-DatasetName", "I-DatasetName", "I-DatasetName", "I-DatasetName" ]
[ "Method:", "SPLAT", "Hyperparameter:", "Mj", "Hyperparameter", "value:", "-", "Metric:", "Joint", "Goal", "Accuracy", "(JGA)", "Metric", "value:", "2021", "Task:", "Named", "Entity", "Recognition", "Dataset:", "Schema-Guided", "Dialogue", "(SGD)" ]
[ "B-MethodName", "I-MethodName", "B-HyperparameterName", "I-HyperparameterName", "B-HyperparameterName", "O", "O", "B-MetricName", "I-MetricName", "I-MetricName", "I-MetricName", "I-MetricName", "B-MetricName", "B-MetricValue", "I-MetricValue", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "B-DatasetName", "I-DatasetName", "I-DatasetName", "I-DatasetName" ]
[ "Method:", "SPLAT", "Hyperparameter:", "IletR", "Hyperparameter", "value:", "-", "Metric:", "Intent", "Accuracy", "Metric", "value:", "2021", "Task:", "Named", "Entity", "Recognition", "Dataset:", "Schema-Guided", "Dialogue", "(SGD)" ]
[ "B-MethodName", "I-MethodName", "B-HyperparameterName", "I-HyperparameterName", "B-HyperparameterName", "O", "O", "B-MetricName", "I-MetricName", "I-MetricName", "B-MetricName", "B-MetricValue", "I-MetricValue", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "B-DatasetName", "I-DatasetName", "I-DatasetName", "I-DatasetName" ]
[ "Method:", "SPLAT", "Hyperparameter:", "M", "Hyperparameter", "value:", "-", "Metric:", "Joint", "Goal", "Accuracy", "(JGA)", "Metric", "value:", "2021", "Task:", "Named", "Entity", "Recognition", "Dataset:", "Schema-Guided", "Dialogue", "(SGD)" ]
[ "B-MethodName", "I-MethodName", "B-HyperparameterName", "I-HyperparameterName", "B-HyperparameterName", "O", "O", "B-MetricName", "I-MetricName", "I-MetricName", "I-MetricName", "I-MetricName", "B-MetricName", "B-MetricValue", "I-MetricValue", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "B-DatasetName", "I-DatasetName", "I-DatasetName", "I-DatasetName" ]
[ "Method:", "SPLAT", "Hyperparameter:", "R1", "Hyperparameter", "value:", "-", "Metric:", "Intent", "Accuracy" ]
[ "B-MethodName", "I-MethodName", "B-HyperparameterName", "I-HyperparameterName", "B-HyperparameterName", "O", "O", "B-MetricName", "I-MetricName", "I-MetricName" ]
[ "keep", "the", "attention", "window", "size", "set", "to", "512", ".", "The", "maximum", "sequence", "length", "is", "4096", ".", "During", "pretraining", "we", "train", "the", "base", "model", "for", "a", "total", "of", "850k", "training", "steps", "and", "the", "large", "model", "for", "800k", "training", "steps", ".", "During", "fineuning", "we", "train", "all", "models", "for", "a", "single", "run", "of", "10", "epochs", "and", "choose", "the", "model", "with", "the", "highest", "joint", "goal", "accuracy", "on", "the", "development", "set", ".", "We", "use", "the", "Adam", "optimizer", "(", "Kingma", "and", "Ba", ",", "2014", ")", "with", "a", "maximum", "learning", "rate", "of", "10−5which", "is", "warmed", "up", "for", "the", "first", "10", "%", "of", "steps", "and", "subsequently", "decays", "linearly", ".", "We", "set", "the", "batch", "size", "to", "32", "for", "base", "models", "and", "to", "16", "for", "large", "models", ".", "We", "pre", "-", "train", "SPLAT", "on", "English", ",", "2022", ")", "BERT", "-", "base", "(", "110", "M", ")", "✗", "-", "56.1", "AG", "-", "DST", "(", "Tian", "et", "al", ".", ",", "2021", ")", "GPT-2", "(", "117", "M", ")", "✗", "‡-", "56.1", "AG", "-", "DST", "(", "Tian", "et", "al", ".", ",", "2021", ")", "PLATO-2", "(", "310", "M", ")", "✗", "‡-", "57.3", "DaP", "(", "seq", ")", "(", "Lee", "et", "al", ".", ",", "2021", ")", "T5", "-", "base", "(", "220", "M", ")", "✓", "-", "51.2", "DaP", "(", "ind", ")", "(", "Lee", "et", "al", ".", ",", "2021", ")", "T5", "-", "base", "(", "220", "M", ")", "✗", "-", "57.5", "D3ST", "(", "Base", ")", "(", "Zhao", "et", "al", ".", ",", "2022", ")", "T5", "-", "base", "(", "220", "M", ")", "✓", "-", "56.1", "D3ST", "Perform", "Entity", "Recognition", "(NER)", "to", "label", "text", "into", "different", "categories.", "Start", "each", "category", "prediction", "in", "a", "new", "line.", "Text:", "focus", "on", "more", "difficult", "exam-", "ples", "and", "is", "trained", "for", "a", "fixed", "schema", ".", "DaP", "(", "ind", ")", "also", "achieves", "similar", "performance", "but", "needs", "one", "inference", "pass", "for", "every", "slot", "at", "every", "turn", "of", "the", "dialogue", ".", "This", "is", "much", "slower", "and", "simply", "not", "real-", "istic", "in", "real", "-", "world", "scenarios", "with", "a", "large", "number", "of", "available", "services", "and", "slots", ".", "The", "sequential", "variant", "DaP", "(", "seq", ")", "which", "instead", "outputs", "the", "full", "state", "in", "a", "single", "pass", "performs", "much", "worse", ".", "Comparison", ".", "While", "DaP", "(", "ind", ")", "shows", "strong", "per-", "formance", "that", "matches", "SPLAT", "on", "MultiWOZ", ",", "SPLAT", "fares", "much", "better", "than", "DaP", "(", "ind", ")", "on", "the", "SGD", "dataset", ".", "This", "can", "be", "seen", "to", "be", "indicative", "of", "a", "stronger", "generalization", "ability", "as", "MultiWOZ", "uses", "the", "same", "schema", "at", "training", "and", "test", "time", "whereas", "SGD", "includes", "new", ",", "unseen", "services", "at", "test", "time", "and", "thus", "requires", "the", "model", "to", "generalize", "and", "under-", "stand", "the", "natural", "language", "schema", "descriptions", ".", "4.3", "Robustness", "DST", "models", "which", "take", "natural", "language", "descrip-", "tions", "of", "intents", "and", "slots", "as", "input", "naturally", "may", "be", "sensitive", "to", "changes", "in", "these", "descriptions", ".", "In", "order", "to", "evaluate", "the", "robustness", "of", "our", "model", "to", "such", "linguistic", "variations", "we", "perform", "experiments", "on", "the", "SGD", "-", "X", "benchmark", ".", "The", "SGD", "-", "X", "benchmark", "comes", "with", "five", "crowd", "-", "sourced", "schema", "variants", "v1", "to", "v5", "which", "increasingly", "diverge", "in", "style", "from", "the", "original", "schema", ".", "We", "train", "SPLAT", "on", "SGD", "and", "evaluate", "it", "on", "the", "test", "set", "using", "all", "five", "different", "schema", "variants", ".", "As", "shown", "in", "Table", "3", ",", "our", "model", "is", "considerably", "more", "robust", "to", "linguistic", "variations", "than", "all", "of", "the", "84", "SGD", "MultiWOZ", "Model", "Params", ".", "Intent", "JGA", "Intent", "JGA", "Longformer", "(", "extr", ".", ")", "110", "M", "95.9", "78.5", "91.4", "55.5", "+", "SPM", "110", "M", "97.0", "79.0", "91.4", "56.1", "+", "SPM", "+", "RSS", "-", "PT", "110", "M", "96.7", "80.1", "91.4", "56.6", "Longformer", "(", "extr", ".", ")", "340", "M", "97.5", "83.5", "91.4", "56.3", "+", "SPM", "340", "M", "98.2", "83.8", "91.4", "57.8", "+", "SPM", "+", "RSS", "-", "PT", "340", "M", "97.6", "85.3", "91.5", "57.4", "Table", "5", ":", "Abl", "categorical", "and", "handcrafted", "features", "for", "categorical", "slots", ".", "Generative", "DST", ".", "In", "an", "attempt", "to", "address", "the", "lack", "of", "ability", "to", "generalize", "to", "new", "domains", "and", "ontologies", ",", "Wu", "et", "al", ".", "(", "2019", ")", "propose", "incorporating", "a", "generative", "component", "into", "DST", ".", "Based", "on", "the", "dialog", "history", "and", "a", "domain", "-", "slot", "pair", "a", "state", "genera-", "tor", "decodes", "a", "value", "for", "each", "slot", ".", "However", "as", "each", "slot", "is", "decoded", "independently", "the", "approach", "can", "not", "model", "slot", "interdependencies", ".", "Feng", "et", "al", ".", "(", "2021", ")", "in-", "stead", "generate", "the", "entire", "state", "as", "a", "single", "sequence", "of", "pointers", "to", "the", "dialogue", "history", "and", "input", "schema", "but", "separately", "encode", "history", "and", "schema", ".", "Zhao", "et", "al", ".", "(", "2021", ")", "model", "DST", "fully", "as85", "a", "string", ".", "Lin", "et", "al", ".", "(", "2021", ")", "transfer", "a", "language", "model", "fine", "-", "tuned", "for", "seq2seq", "Association", "for", "Computational", "Linguis-", "tics.86", "Raghav", "Gupta", ",", "Harrison", "Lee", ",", "Jeffrey", "Zhao", ",", "Yuan", "Cao", ",", "Abhinav", "Rastogi", ",", "and", "Yonghui", "Wu", ".", "2022", ".", "Show", ",", "do", "n’t", "tell", ":", "Demonstrations", "outperform", "descriptions", "for", "schema", "-", "guided", "task", "-", "oriented", "dialogue", ".", "In", ".", "Ting", "Han", ",", "Ximing", "Liu", ",", "Ryuichi", "Takanabu", ",", "Yixin", "Lian", ",", "Chongxuan", "Huang", ",", "Dazhen", "W", ",", "and", ".", "2022", ".", "LUNA:87", "Learning", "slot", "-", "turn", "alignment", "for", "dialogue", "state", "track-", "ing", ".", "In", "Proceedings", "of", "the", "2022", "Conference", "of", "the", ",", "pages", "3319–3328", ",", ",", "Deepak", "Ramachandran", ",", "and", "Alan", "Black", ".", "2013", ".", "The", "dialog", "state", "tracking", "challenge", ".", "In", "Proceedings", "of", "the", "SIGDIAL", "2013", "Conference", ",", "pages", "404–413", ",", ".", "Association", "for", "Computational", "Linguistics", "." ]
[ "O", "O", "B-HyperparameterName", "B-HyperparameterName", "B-HyperparameterName", "O", "O", "B-HyperparameterValue", "O", "O", "B-HyperparameterName", "B-HyperparameterName", "B-HyperparameterName", "O", "B-HyperparameterValue", "O", "O", "B-HyperparameterName", "O", "B-HyperparameterName", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterValue", "I-HyperparameterName", "I-HyperparameterName", "O", "O", "B-HyperparameterValue", "I-HyperparameterValue", "O", "B-HyperparameterValue", "I-HyperparameterName", "I-HyperparameterName", "O", "O", "B-HyperparameterName", "O", "B-HyperparameterName", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterValue", "I-HyperparameterValue", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "B-MetricName", "B-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "B-MethodName", "O", "B-MethodName", "O", "B-MethodName", "O", "O", "O", "B-HyperparameterName", "B-HyperparameterName", "B-HyperparameterName", "O", "B-HyperparameterValue", "O", "O", "O", "O", "O", "O", "B-HyperparameterValue", "O", "O", "B-HyperparameterName", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterName", "I-HyperparameterName", "O", "B-HyperparameterValue", "O", "B-HyperparameterValue", "O", "O", "O", "B-HyperparameterValue", "O", "B-HyperparameterValue", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "B-DatasetName", "O", "O", "O", "B-MethodName", "O", "I-MethodName", "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", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "B-MethodName", "O", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "B-MethodName", "O", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "I-MethodName", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "B-MethodName", "O", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "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", "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", "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", "B-MethodName", "O", "B-DatasetName", "O", "B-MethodName", "O", "O", "O", "O", "B-MethodName", "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", "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", "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", "B-DatasetName", "O", "B-DatasetName", "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", "O", "B-MethodName", "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", "B-DatasetName", "B-DatasetName", "O", "O", "O", "B-MetricName", "I-MetricName", "B-MetricName", "I-MetricName", "B-MethodName", "O", "O", "O", "O", "B-HyperparameterValue", "B-HyperparameterName", "B-MetricValue", "I-MetricValue", "B-MetricValue", "I-MetricValue", "O", "B-MethodName", "B-HyperparameterValue", "B-HyperparameterName", "B-MetricValue", "I-MetricValue", "B-MetricValue", "I-MetricValue", "O", "B-MethodName", "O", "O", "O", "O", "B-HyperparameterValue", "B-HyperparameterName", "B-MetricValue", "I-MetricValue", "B-MetricValue", "I-MetricValue", "B-MethodName", "O", "O", "O", "O", "B-HyperparameterValue", "B-HyperparameterName", "B-MetricValue", "I-MetricValue", "B-MetricValue", "I-MetricValue", "O", "B-MethodName", "B-HyperparameterValue", "B-HyperparameterName", "B-MetricValue", "I-MetricValue", "B-MetricValue", "I-MetricValue", "O", "B-MethodName", "O", "O", "O", "O", "B-HyperparameterValue", "B-HyperparameterName", "B-MetricValue", "I-MetricValue", "B-MetricValue", "I-MetricValue", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "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", "O", "O", "B-MethodName", "O", "O", "O", "B-TaskName", "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", "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", "B-TaskName", "B-TaskName", "O" ]
[ "Proceedings", "of", "the", "61st", "Annual", "Meeting", "of", "the", "Association", "for", "Computational", "Linguistics", "Volume", "1", ":", "Long", "Papers", ",", "pages", "338–350", "July", "9", "-", "14", ",", "2023", "©", "2023", "Association", "for", "Computational", "Linguistics", "Text", "Adversarial", "Purification", "as", "Defense", "against", "Adversarial", "Attacks", "Linyang", "Li", "Demin", "Song", ",", "Xipeng", "Qiu", "School", "of", "Computer", "Science", ",", "Fudan", "University", "Shanghai", "Key", "Laboratory", "of", "Intelligent", "Information", "Processing", ",", "Fudan", "University", "{", "linyangli19", ",", "dmsong20", ",", "xpqiu", "}", "@", "fudan.edu.cn", "Abstract", "Adversarial", "purification", "is", "a", "successful", "defense", "mechanism", "against", "adversarial", "attacks", "without", "requiring", "knowledge", "of", "the", "form", "of", "the", "incom-", "ing", "attack", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "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", "B-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Generally", ",", "adversarial", "purification", "aims", "to", "remove", "the", "adversarial", "perturbations", "therefore", "can", "make", "correct", "predictions", "based", "on", "the", "recovered", "clean", "samples", "." ]
[ "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Despite", "the", "success", "of", "adversarial", "purification", "in", "the", "com-", "puter", "vision", "field", "that", "incorporates", "generative", "models", "such", "as", "energy", "-", "based", "models", "and", "dif-", "fusion", "models", ",", "using", "purification", "as", "a", "defense", "strategy", "against", "textual", "adversarial", "attacks", "is", "rarely", "explored", "." ]
[ "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" ]
[ "In", "this", "work", ",", "we", "introduce", "a", "novel", "adversarial", "purification", "method", "that", "fo-", "cuses", "on", "defending", "against", "textual", "adversarial", "attacks", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "O" ]
[ "With", "the", "help", "of", "language", "models", ",", "we", "can", "inject", "noise", "by", "masking", "input", "texts", "and", "reconstructing", "the", "masked", "texts", "based", "on", "the", "masked", "language", "models", "." ]
[ "O", "O", "O", "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", "this", "way", ",", "we", "con-", "struct", "an", "adversarial", "purification", "process", "for", "textual", "models", "against", "the", "most", "widely", "used", "word", "-", "substitution", "adversarial", "attacks", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "O" ]
[ "We", "test", "our", "proposed", "adversarial", "purification", "method", "on", "several", "strong", "adversarial", "attack", "methods", "in-", "cluding", "Textfooler", "and", "BERT", "-", "Attack", "and", "exper-", "imental", "results", "indicate", "that", "the", "purification", "al-", "gorithm", "can", "successfully", "defend", "against", "strong", "word", "-", "substitution", "attacks", "." ]
[ "O", "O", "O", "O", "B-MethodName", "I-MethodName", "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", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "O" ]
[ "1", "Introduction", "Adversarial", "examples", "(", "Goodfellow", "et", "al", ".", ",", "2014", ")", "can", "successfully", "mislead", "strong", "neural", "models", "in", "both", "computer", "vision", "tasks", "(", "Carlini", "and", "Wagner", ",", "2016", ")", "and", "language", "understanding", "tasks", "(", "Alzantot", "et", "al", ".", ",", "2018", ";", "Jin", "et", "al", ".", ",", "2019", ")", "." ]
[ "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", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "An", "adversarial", "example", "is", "a", "maliciously", "crafted", "example", "attached", "with", "an", "imperceptible", "perturbation", "and", "can", "mislead", "neural", "networks", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "To", "defend", "attack", "examples", "of", "images", ",", "the", "most", "effective", "method", "is", "adversarial", "training", "(", "Goodfellow", "et", "al", ".", ",", "2014", ";", "Madry", "et", "al", ".", ",", "2019", ")", "which", "is", "a", "mini", "-", "max", "game", "used", "to", "incorporate", "perturba-", "tions", "into", "the", "training", "process", ".", "Defending", "adversarial", "attacks", "is", "extremely", "im-", "portant", "in", "improving", "model", "robustness", "." ]
[ "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", "B-TaskName", "I-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "How-", "ever", ",", "defending", "adversarial", "examples", "in", "natural", "languages", "is", "more", "challenging", "due", "to", "the", "discrete", "nature", "of", "texts", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "That", "is", ",", "gradients", "can", "not", "be", "used", "directly", "in", "crafting", "perturbations", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "The", "substitution-", "based", "adversarial", "examples", "are", "more", "complicated", "than", "gradient", "-", "based", "adversarial", "examples", "in", "im-", "ages", ",", "making", "it", "difficult", "for", "neural", "networks", "to", "de-", "fend", "against", "these", "substitution", "-", "based", "attacks", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "The", "first", "challenge", "of", "defending", "against", "adver-", "sarial", "attacks", "in", "NLP", "is", "that", "due", "to", "the", "discrete", "na-", "ture", ",", "these", "substitution", "-", "based", "adversarial", "examples", "can", "have", "substitutes", "in", "any", "token", "of", "the", "sentence", "and", "each", "substitute", "has", "a", "large", "candidate", "list", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "This", "would", "cause", "a", "combinatorial", "explosion", "problem", ",", "making", "it", "hard", "to", "apply", "adversarial", "training", "meth-", "ods", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Strong", "attacking", "methods", "such", "as", "Jin", "et", "al", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "(", "2019", ")", "show", "that", "using", "the", "crafted", "adversarial", "ex-", "amples", "as", "data", "augmentation", "in", "adversarial", "training", "can", "not", "effectively", "defend", "against", "these", "substitution-", "based", "attacks", "." ]
[ "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" ]
[ "Further", ",", "defending", "strategies", "such", "as", "adversarial", "training", "rely", "on", "the", "assumption", "that", "the", "candidate", "lists", "of", "the", "substitutions", "are", "accessible", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "However", ",", "the", "candidate", "lists", "of", "the", "substitutions", "should", "notbe", "exposed", "to", "the", "target", "model", ";", "that", "is", ",", "the", "target", "model", "should", "be", "unfamiliar", "to", "the", "candi-", "date", "list", "of", "the", "adversarial", "examples", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "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", "real", "-", "world", "defense", "systems", ",", "the", "defender", "is", "not", "aware", "of", "the", "strategy", "the", "potential", "attacks", "might", "use", ",", "so", "the", "as-", "sumption", "that", "the", "candidate", "list", "is", "available", "would", "significantly", "constrain", "the", "potential", "applications", "of", "these", "defending", "methods", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Considering", "that", "it", "is", "challenging", "to", "defend", "against", "textual", "adversarial", "attacks", "when", "the", "form", "of", "the", "attacks", "can", "not", "be", "acknowledged", "in", "advance", ",", "we", "introduce", "a", "novel", "adversarial", "purification", "method", "as", "a", "feasible", "defense", "mechanism", "against", "these", "at-", "tacks", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "The", "adversarial", "purification", "method", "is", "to", "purify", "adversarially", "perturbed", "input", "samples", "before", "making", "predictions", "(", "Srinivasan", "et", "al", ".", ",", "2021", ";", "Shi338", "et", "al", ".", ",", "2021", ";", "Yoon", "et", "al", ".", ",", "2021", ")", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "The", "major", "works", "about", "adversarial", "purification", "focus", "on", "purifying", "continuous", "inputs", "such", "as", "images", ",", "therefore", "these", "works", "explore", "different", "generative", "models", "such", "as", "GANs", "(", "Samangouei", "et", "al", ".", ",", "2018", ")", ",", "energy", "-", "based", "models", "(", "EBMs", ")", "(", "LeCun", "et", "al", ".", ",", "2006", ")", "and", "recently", "developed", "diffusion", "models", "(", "Song", "et", "al", ".", ",", "2021", ";", "Nie", "et", "al", ".", ",", "2022", ")", "." ]
[ "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-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" ]
[ "However", ",", "in", "textual", "adversarial", "at-", "tacks", ",", "the", "inputs", "are", "discrete", "tokens", "which", "makes", "it", "more", "challenging", "to", "deploy", "previous", "adversarial", "purification", "methods", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Therefore", ",", "we", "introduce", "a", "purification", "mecha-", "nism", "with", "the", "help", "of", "masked", "language", "models", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "We", "first", "consider", "the", "widely", "used", "masking", "process", "to", "inject", "noise", "into", "the", "input", ";", "then", "we", "recover", "the", "clean", "texts", "from", "the", "noisy", "inputs", "with", "the", "help", "of", "the", "masked", "language", "models", "(", "e.g", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "a", "BERT", "(", "De-", "vlin", "et", "al", ".", ",", "2018", ")", ")", "." ]
[ "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Further", ",", "considering", "that", "the", "iterative", "process", "in", "previous", "adversarial", "purifica-", "tion", "algorithms", "can", "be", "extremely", "costly", "(", "e.g", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "a", "VP", "-", "SDE", "process", "in", "diffusion", "models", "(", "Song", "et", "al", ".", ",", "2021", ")", ")", ",", "we", "instead", "simplify", "the", "iterative", "process", "to", "an", "ensemble", "-", "purifying", "process", "that", "conducting", "adversarial", "purification", "multiple", "times", "to", "obtain", "an", "ensembled", "result", "as", "a", "compromise", "to", "the", "time", "cost", "in", "traditional", "adversarial", "purification", "process", "." ]
[ "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" ]
[ "Through", "extensive", "experiments", ",", "we", "prove", "that", "the", "proposed", "text", "adversarial", "purification", "algorithm", "can", "successfully", "serve", "as", "defense", "against", "strong", "at-", "tacks", "such", "as", "Textfooler", "and", "BERT", "-", "Attack", "." ]
[ "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", "I-TaskName", "O" ]
[ "Exper-", "iment", "results", "show", "that", "the", "accuracy", "under", "attack", "in", "baseline", "defense", "methods", "is", "lower", "than", "random", "guesses", ",", "while", "after", "text", "purification", ",", "the", "perfor-", "mance", "can", "reach", "only", "a", "few", "percent", "lower", "than", "the", "original", "accuracy", "when", "the", "candidate", "range", "of", "the", "attack", "is", "limited", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Further", ",", "extensive", "results", "indicate", "that", "the", "candidate", "range", "of", "the", "attacker", "score", "is", "essential", "for", "successful", "attacks", ",", "which", "is", "a", "key", "fac-", "tor", "in", "maintaining", "the", "semantics", "of", "the", "adversaries", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Therefore", "we", "also", "recommend", "that", "future", "attacking", "methods", "can", "focus", "on", "achieving", "successful", "attacks", "with", "tighter", "constraints", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "To", "summarize", "our", "contributions", ":", "(", "1", ")", "We", "raise", "the", "concern", "of", "defending", "substitution", "-", "based", "adversarial", "attacks", "without", "ac-", "knowledging", "the", "form", "of", "the", "attacks", "in", "NLP", "tasks", "." ]
[ "O", "O", "O", "O", "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" ]
[ "(", "2", ")", "To", "the", "best", "of", "our", "knowledge", ",", "we", "are", "the", "first", "to", "consider", "adversarial", "purification", "as", "a", "defense", "against", "textual", "adversarial", "attacks", "exemplified", "by", "strong", "word", "-", "substitution", "attacks", "and", "combine", "text", "adversarial", "purification", "with", "pre", "-", "trained", "models.", "(", "3", ")", "We", "perform", "extensive", "experiments", "to", "demon-", "strate", "that", "the", "adversarial", "purification", "method", "is", "capable", "of", "defending", "strong", "adversarial", "attacks", ",", "which", "brings", "a", "new", "perspective", "to", "defending", "tex-", "tual", "adversarial", "attacks", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "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" ]
[ "2", "Related", "Work", "2.1", "Adversarial", "Attacks", "in", "NLP", "In", "NLP", "tasks", ",", "current", "methods", "use", "substitution-", "based", "strategies", "(", "Alzantot", "et", "al", ".", ",", "2018", ";", "Jin", "et", "al", ".", ",", "2019", ";", "Ren", "et", "al", ".", ",", "2019", ")", "to", "craft", "adversarial", "exam-", "ples", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "B-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", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Most", "works", "focus", "on", "the", "score", "-", "based", "black-", "box", "attack", ",", "that", "is", ",", "attacking", "methods", "know", "the", "logits", "of", "the", "output", "prediction", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "These", "methods", "use", "different", "strategies", "(", "Yoo", "et", "al", ".", ",", "2020", ";", "Morris", "et", "al", ".", ",", "2020b", ")", "to", "find", "words", "to", "replace", ",", "such", "as", "genetic", "algorithm", "(", "Alzantot", "et", "al", ".", ",", "2018", ")", ",", "greedy", "-", "search", "(", "Jin", "et", "al", ".", ",", "2019", ";", "Li", "et", "al", ".", ",", "2020", ")", "or", "gradient", "-", "based", "methods", "(", "Ebrahimi", "et", "al", ".", ",", "2017", ";", "Cheng", "et", "al", ".", ",", "2019", ")", "and", "get", "substitutes", "using", "synonyms", "(", "Jin", "et", "al", ".", ",", "2019", ";", "Mrkši", "´", "c", "et", "al", ".", ",", "2016", ";", "Ren", "et", "al", ".", ",", "2019", ")", "or", "language", "models", "(", "Li", "et", "al", ".", ",", "2020", ";", "Garg", "and", "Ramakrishnan", ",", "2020", ";", "Shi", "et", "al", ".", ",", "2019", ")", "." ]
[ "O", "O", "O", "O", "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", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-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", "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" ]
[ "2.2", "Adversarial", "Defenses", "We", "divide", "the", "defense", "methods", "for", "word-", "substitution", "attacks", "by", "whether", "the", "defense", "method", "requires", "knowledge", "of", "the", "form", "of", "the", "attack", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "When", "the", "candidate", "list", "is", "known", ",", "recent", "works", "introduce", "defense", "strategies", "that", "incorporate", "the", "candidates", "of", "the", "words", "to", "be", "replaced", "as", "an", "aug-", "mentation", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Jin", "et", "al", "." ]
[ "O", "O", "O", "O" ]
[ "(", "2019", ")", ";", "Li", "et", "al", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "(", "2020", ")", ";", "Si", "et", "al", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "(", "2020", ")", "uses", "generated", "adversaries", "to", "augment", "the", "classifier", "for", "better", "defense", "performances", ";", "Jia", "et", "al", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "(", "2019", ")", ";", "Huang", "et", "al", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "(", "2019", ")", "introduce", "a", "certified", "robust", "model", "to", "construct", "a", "certified", "space", "within", "the", "range", "of", "a", "candidate", "list", "therefore", "the", "substitu-", "tions", "in", "the", "candidate", "list", "can", "not", "perturb", "the", "model", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Zhou", "et", "al", "." ]
[ "O", "O", "O", "O" ]
[ "(", "2020", ")", ";", "Dong", "et", "al", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "(", "2021", ")", "construct", "a", "convex", "hull", "based", "on", "the", "candidate", "list", "which", "can", "resist", "substitutions", "in", "the", "candidate", "list", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "To", "defend", "unknown", "attacks", ",", "NLP", "models", "can", "in-", "corporate", "gradient", "-", "based", "adversarial", "training", "strate-", "gies", "(", "Miyato", "et", "al", ".", ",", "2016", ";", "Madry", "et", "al", ".", ",", "2019", ")", "since", "recent", "works", "(", "Ebrahimi", "et", "al", ".", ",", "2017", ";", "Cheng", "et", "al", ".", ",", "2019", ";", "Zhu", "et", "al", ".", ",", "2019", ";", "Li", "and", "Qiu", ",", "2020", ")", "show", "that", "gradient", "-", "based", "adversarial", "training", "can", "also", "improve", "defense", "performances", "against", "word-", "substitution", "attacks.339", "Adversarial", "Image", "(", "Gibbon", ")", "Purified", "Image", "(", "Panda", ")", "a", "continuous", "image", "re", "-", "construct", "process", "from", "a", "noisy", "image", "e.g", "." ]
[ "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" ]
[ "using", "Diffusion", "Models", "/", "Score", "-", "based", "Models", "I", "possess", "the", "good", "common", "logical", "sense", "to", "realize", "..." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "it", "is", "something", "I", "like", "..." ]
[ "O", "O", "O", "O", "O", "O" ]
[ "Adversarial", "Text", "(", "Negative", ")", "..." ]
[ "O", "O", "O", "O", "O", "O" ]
[ "I", "[", "MASK", "]", "the", "good", "common", "logical", "sense", "to", "[", "MASK", "]", "..." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "it", "is", "something", "I", "like", "..." ]
[ "O", "O", "O", "O", "O", "O" ]
[ "I", "possess", "the", "good", "common", "logical", "[", "MASK", "]", "to", "realize", "..." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "it", "is", "[", "MASK", "]", "I", "like", "..." ]
[ "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "..." ]
[ "O" ]
[ "..." ]
[ "O" ]
[ "MLMI", "know", "the", "good", "common", "logical", "sense", "to", "sense", "..." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "it", "is", "something", "I", "like", "..." ]
[ "O", "O", "O", "O", "O", "O" ]
[ "Purified", "TextsI", "possess", "the", "good", "common", "logical", "feel", "to", "realize", "..." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "it", "is", "something", "I", "like", "..." ]
[ "O", "O", "O", "O", "O", "O" ]
[ "(", "Positive", ")", "I", "possess", "the", "good", "common", "logical", "[", "MASK", "]", "to", "[", "MASK", "]", "..." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]

Dataset Card for "annotation2_wo_elimination"

More Information needed

Downloads last month
0
Edit dataset card