Datasets:

havens2

/

naacl2022

like 0

[ "In", "terms", "of", "Reuters-21678,", "15,", "5,", "and", "11", "disjoint", "classes", "are", "used", "for", "meta-training/validation/test", "sets", "and", "37,", "10,", "and", "24", "disjoint", "classes", "for", "RCV-1." ]

[ 0, 0, 0, 13, 5, 6, 6, 6, 0, 0, 0, 0, 0, 3, 0, 0, 5, 6, 6, 6, 0, 0, 0, 13 ]

[ "Hence,", "the", "Huffpost", "headline", "is", "divided", "into", "20,", "5,", "and", "16", "disjoint", "classes", "for", "meta-training,", "validation,", "and", "test", "sets." ]

[ 0, 0, 13, 14, 0, 0, 0, 5, 6, 6, 6, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "In", "this", "work,", "we", "used", "the", "same", "split", "of", "classes", "as", "in", "(Bao", "et", "al.,", "2019)", "for", "the", "Huffpost", "headline(Misra", "and", "Grover,", "2021),", "Reuters-21578(Lewis.,", "1997),", "and", "RCV-1(Lewis", "et", "al.,", "2004)", "datasets." ]

[ 0, 0, 0, 0, 0, 0, 0, 3, 4, 4, 0, 0, 0, 0, 0, 0, 0, 0, 13, 14, 0, 0, 0, 13, 0, 0, 13, 0, 0, 0, 0 ]

[ "To", "train", "and", "evaluate", "the", "models,", "we", "divided", "each", "of", "the", "aforementioned", "datasets", "into", "a", "metaval),", "and", "training", "set", "(", "S", "test)", "as", "disjoint", "sets", "of", "classes", "meta-test", "set", "(", "within", "the", "experimental", "setting." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "Our", "goal", "is", "to", "match", "reviews", "to", "their", "own", "corresponding", "product", "categories." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "Amazon", "data", "is", "a", "real-world", "dataset", "collected", "from", "Amazon.com", "as", "a", "set", "of", "customer", "reviews", "from", "24", "types", "of", "product", "categories(He", "and", "McAuley,", "2016)." ]

[ 13, 14, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "RCV-1", "is", "a", "set", "of", "newswire", "stories", "published", "by", "Reuters", "journalists", "from", "1996", "to", "1997", "(Lewis", "et", "al.,", "2004)", "and", "comprises", "71", "topic", "classes." ]

[ 13, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "In", "addition,", "we", "use", "the", "ApteMod", "version", "and", "discard", "documents", "with", "more", "than", "one", "label", "to", "avoid", "ambiguity,", "and", "thus", "31", "classes", "remain." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "Reuters-21578", "is", "composed", "of", "documents", "that", "appeared", "on", "the", "Reuters", "newswire", "in", "1987", "(Lewis.,", "1997)." ]

[ 13, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "It", "is", "composed", "of", "41", "topics." ]

[ 0, 0, 0, 0, 0, 0 ]

[ "Huffpost", "Headlines", "is", "a", "collection", "of", "news", "headlines", "published", "in", "the", "Huffington", "Post", "from", "2012", "to", "2018", "(Misra", "and", "Grover,", "2021)." ]

[ 13, 14, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "newsgroup", "posts", "for", "20", "topics", "(Lang,", "1995)." ]

[ 0, 0, 0, 0, 0, 0, 0 ]

[ "20", "Newsgroups", "is", "a", "collection", "of", "discourses", "in" ]

[ 13, 14, 0, 0, 0, 0, 0, 0 ]

[ "We", "introduce", "the", "datasets", "and", "the", "split", "(i.e.,", "train/val/test)", "which", "had", "been", "maintained", "in", "our", "experiments." ]

[ 0, 0, 0, 0, 0, 0, 3, 0, 3, 0, 0, 0, 0, 0, 0, 0 ]

[ "A.1", "Datasets" ]

[ 0, 0 ]

[ "A", "Appendix" ]

[ 0, 0 ]

[ "References" ]

[ 0 ]

[ "As", "a", "result,", "we", "proposed", "a", "novel", "meta-learning", "framework", "for", "the", "learning-to-attend", "and", "showed", "that", "LEA", "is", "an", "effective", "method", "that", "facilitates", "the", "utilization", "of", "large-sized", "PTMs", "in", "few-shot", "learning." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 11, 12 ]

[ "We", "have", "attempted", "to", "design", "a", "novel", "embedding", "transfer", "method", "for", "deriving", "a", "meta-level", "attention", "aspects", "dictionary", "to", "enable", "a", "new", "task", "to", "simply", "borrow", "the", "most", "relevant", "attention", "aspects", "from", "the", "dictionary." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "We", "hypothesized", "that", "a", "type", "of", "task-specific", "selfattentive", "mechanism", "might", "improve", "few-shot", "learning", "performance,", "especially", "when", "it", "is", "prohibitive", "to", "fine-tune", "a", "large-sized", "PTM." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "This", "empirical", "result", "indicates", "that", "each", "task", "derives", "its", "optimal", "document", "embedding" ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "As", "shown", "in", "Figure", "2,", "all", "the", "datasets", "exhibit", "their", "best", "performance", "when", "setting", "the", "top-k", "attention", "aspects", "to", "20." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3, 4, 4, 0, 5 ]

[ "We", "fixed", "the", "size", "of", "the", "meta", "attention", "aspects", "dictionary", "to", "150", "and", "measured", "the", "performances", "by", "gradually", "scaling", "the", "k", "up", "to", "1,", "10,", "20,", "30,", "50,", "75,", "150." ]

[ 0, 0, 0, 3, 4, 4, 4, 4, 4, 4, 0, 5, 0, 0, 0, 0, 0, 0, 0, 0, 3, 0, 0, 5, 5, 5, 5, 5, 5, 5 ]

[ "This", "specific", "study", "was", "conducted", "on", "the", "same", "frozen", "BERTBASE", "as", "the", "underlying", "PTM", "with", "the", "5-way", "5-shot", "experiment", "for", "the", "all", "datasets." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 11, 12, 0, 0, 0, 0, 0 ]

[ "We", "also", "investigate", "the", "impact", "of", "the", "number", "(i.e.,", "k)", "of", "task-specific", "attention", "aspects." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 3, 0, 0, 0, 0 ]

[ "6", "Conclusion" ]

[ 0, 0 ]

[ "Number", "of", "Top-k", "Attention", "Aspects" ]

[ 0, 0, 0, 0, 0 ]

[ "5.3", "Hyperparameter", "Study:", "Effect", "of", "the" ]

[ 0, 0, 0, 0, 0, 0 ]

[ "These", "plots", "demonstrate", "that", "LEA", "produces", "a", "structured", "task-specific", "embedding", "space", "after", "our", "task-adaptation", "step." ]

[ 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "Figure", "3c", "shows", "the", "relationships", "between", "the", "four", "classes", "of", "the", "‘science’", "domain", "and", "the", "others", "on", "the", "space." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "Figure", "3b", "shows", "the", "relationships", "between", "the", "‘recreation’", "domain", "composed", "of", "four", "classes", "and", "the", "rest", "on", "the", "space." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "To", "qualitatively", "characterize", "the", "task-specific", "document", "embedding", "space,", "we", "split", "20", "Newsgroup", "into", "seven", "top-level", "domains,", "that", "is,", "‘atheism’,", "‘computer’,", "‘for-sale’,", "‘recreation’,", "‘science’,", "‘religion’,", "and", "‘talk’", "and", "projected", "them", "via", "t-SNE", "as", "shown", "in", "Figure", "3a." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 13, 14, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "In", "addition,", "we", "plots", "the", "task-specific", "document", "embeddings", "and", "observe", "the", "relationships", "among", "classes", "on", "20", "Newsgroups", "dataset." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 13, 14, 0 ]

[ "Visualization" ]

[ 0 ]

[ "5.4", "Task-Specific", "Document", "Embedding" ]

[ 0, 0, 0, 0 ]

[ "by", "referring", "only", "to", "the", "most", "relevant", "subset", "rather", "than", "exploiting", "all", "meta-level", "attention", "aspects." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "Namely,", "the", "results", "demonstrate", "that", "LEA", "quickly", "recognizes", "how", "to", "attend", "for", "new", "tasks", "using", "the", "established", "meta-attention", "aspects", "and", "provides", "a", "robust", "performance", "in", "few-shot", "text", "classification", "problems." ]

[ 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 11, 12, 12, 0 ]

[ "As", "shown", "in", "Table", "1", "and", "2,", "LEA", "exhibits", "the", "competitive", "performance", "in", "both", "the", "5-way", "1-shot", "and", "5-way", "5-shot,", "compared", "to", "the", "state-of-the-arts", "for", "all", "the", "datasets." ]

[ 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 11, 12, 0, 11, 12, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "All", "performance", "scores", "are", "reported", "as", "the", "average", "for", "three", "repetitions." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "We", "additionally", "applied", "LEA", "on", "RoBERTaBASE", "(Liu", "et", "al.,", "2019)", "and", "fastText", "(Bojanowski", "et", "al.,", "2017)", "to", "verify", "the", "applicability", "of", "LEA." ]

[ 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1 ]

[ "For", "the", "comparison", "with", "Frog-GNN,", "we", "referred", "to", "the", "reported", "results", "from", "(Xu", "and", "Xiang,", "2021)." ]

[ 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "As", "in", "LEA,", "DS", "is", "given", "the", "[CLS]", "embedding", "and", "the", "token", "embeddings", "from", "BERT’s", "last", "layer,", "whereas", "the", "other", "algorithms", "used", "the", "[CLS]", "embedding", "of", "BERT." ]

[ 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "5.2", "Overall", "Performance" ]

[ 0, 0, 0 ]

[ "Here,", "MAML", "(Finn", "et", "al.,", "2017)", "denotes", "the", "representative", "optimization-based", "meta-learning", "algorithm,", "PROTO", "(Snell", "et", "al.,", "2017)", "indicates", "the", "prototype", "network,", "LEO", "(Rusu", "et", "al.,", "2018)", "denotes", "the", "meta-learning", "algorithm", "using", "latent", "embedding", "optimization,", "INDUCTION", "indicates", "the", "induction", "network", "(Geng", "et", "al.,", "2019),", "DS", "(Bao", "et", "al.,", "2019)", "denotes", "few-shot", "text", "classification", "algorithm", "using", "the", "underlying", "word", "distributions,", "and", "Frog-GNN", "(Xu", "and", "Xiang,", "2021)", "denotes", "the", "multi-perspective", "aggregation", "based", "graph", "neural", "network." ]

[ 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 2, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 2, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "MAML", "alternates", "between", "two", "update", "processes", "during", "meta-training:", "(1)", "task-adaptation", "and", "(2)", "meta-optimization." ]

[ 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "For", "a", "parametric", "model", "fθ,", "MAML", "seeks", "to", "find", "task-specific", "parameters", "θi", "for", "any", "new", "task", "τi", "sampled", "from", "a", "particular", "disp(τ", "),", "tribution", "of", "tasks." ]

[ 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "In", "this", "experiment,", "we", "evaluate", "and", "compare", "LEA", "with", "six", "state-of-art", "methods", "as", "follows:" ]

[ 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0 ]

[ "5.1", "Baselines" ]

[ 0, 0 ]

[ "The", "details", "of", "the", "datasets", "are", "introduced", "in", "Appendix", "A.1." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "We", "conducted", "two", "different", "experiments:", "5-way", "1-shot", "and", "5-way", "5shot", "all", "over", "datasets." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "We", "evaluated", "LEA", "on", "five", "text", "datasets", "—", "20", "Newsgroup(Lang,", "1995),", "Huffpost", "headline(Misra", "and", "Grover,", "2021),", "Reuters-21578(Lewis.,", "1997),", "RCV1(Lewis", "et", "al.,", "2004),", "and", "Amazon", "product", "reviews", "(He", "and", "McAuley,", "2016)", "—", "and", "compared", "it", "with", "current", "state-of-art", "methods." ]

[ 0, 0, 1, 0, 0, 0, 0, 0, 13, 14, 0, 13, 14, 0, 0, 0, 13, 0, 13, 0, 0, 0, 0, 13, 14, 14, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "5", "Experimental", "Results" ]

[ 0, 0, 0 ]

[ "During", "the", "meta-optimization", "step,", "the", "groups", "are", "." ]

[ 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "The", "former", "proceeds", "as", "follows:" ]

[ 0, 0, 0, 0, 0 ]

[ "As", "noted", "in", "Algorithm", "1,", "LEA", "alternates", "the", "following", "two", "update", "steps:", "(1)", "task", "adaptation", "(or", "inner-update)", "and", "(2)", "meta-optimization", "(or", "outerupdate)." ]

[ 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "4.4", "Meta-Training", "Objectives" ]

[ 0, 0, 0 ]

[ "E", "L", "is", "the", "self-attentive", "document", "where", "embedding", "of", "the", "jth", "input", "of", "the", "task", "τi,", "and", "H", "τi", "L", "is", "a", "set", "of", "token", "embedding", "vectors", "for", "the", "jth", "instance", "with", "L", "tokens", "in", "the", "task", "τi." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "This", "is", "formulated", "as", "follows:" ]

[ 0, 0, 0, 0, 0 ]

[ "For", "a", "text", "input,", "we", "utilize", "the", "corresponding", "embedding", "vectors", "for", "the", "individual", "tokens,", "which", "are", "denoted", "as", "H", "τi", "j,L]", "for", "the", "jth", "text", "example", "of", "the", "task", "τi." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "We", "then", "apply", "it", "into", "the", "generation", "of", "document", "embeddings", "for", "text" ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "Here,", "we", "perform", "the", "self-attentive", "feature", "extraction", "using", "the", "aforementioned", "top-k", "task-specific", "attention", "aspects", "for", "a", "task." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "4.3", "Task-Specific", "Self-Attentive", "Document" ]

[ 0, 0, 0, 0 ]

[ "This", "is", "formulated", "as", "follows:" ]

[ 0, 0, 0, 0, 0 ]

[ "As", "a", "result,", "we", "can", "extract", "the", "top-k", "task-specific", "attention", "aspects", "for", "the", "task", "τi." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "The", "gating", "process", "G", "produces", "a", "sparse", "output", "vector", "by", "beAN", ",", "Wn", "∈", "ing", "parameterized", "with", "RAN", "×", ",", "where", "the", "remaining", "values", "except", "for", "the", "k", "elements", "are", "forced", "to", "become", "zeros,", "and", "the", "top-k", "weights", "are", "finally", "generated", "through", "a", "softmax", "function." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "where", "gτi", "n", "is", "the", "gating", "output", "vector", "whose", "number", "of", "dimensions", "must", "be", "the", "same", "as", "the", "size", "of", "the", "meta-attention-aspects", "dictionary." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "The", "gating", "output", "vector", "is", "calculated", "through", "the", "following", "formulation:" ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "Using", "the", "aforementioned", "class", "embedding,", "we", "attempt", "to", "selectively", "(i.e.,", "top-k)", "collect", "taskspecific", "attention", "aspects", "for", "a", "given", "task", "by", "employing", "a", "gating", "mechanism", "(Shazeer", "et", "al.,", "2017)." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "As", "a", "result,", "the", "class", "embedding", "eτi", "n", "is", "enforced", "to", "encode", "the", "pairwise", "relationship", "with", "other", "classes." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "[CLS]", "(denoted", "as", "cτi", "j", "for", "the", "case", "of", "the", "jth", "text", "instance", "of", "a", "specific", "task", "τi)", "of", "a", "text", "instance", "in", "PTMs", "(Devlin", "et", "al.,", "2019;", "Lan", "et", "al.,", "2019;", "Liu", "et", "al.,", "2019)." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "(cid:0)", "(3)", "where", "eτi", "n", "is", "the", "representative", "embedding", "for", "the", "particular", "class", "n", "under", "a", "given", "task", "τi,", "fθr", "indicates", "the", "relation", "network", "(Sung", "et", "al.,", "2018),", "and", "fθe", "is", "an", "encoder", "network", "that", "transforms", "the", "delegate", "embedding" ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "First,", "each", "task", "is", "fed", "into", "an", "encoding", "process,", "which", "is", "formulated", "as", "follows:" ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "To", "do", "so,", "we", "assess", "the", "relevance", "of", "the", "task", "among", "the", "meta-level", "attention", "aspects", "WA." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "Note", "that", "k", "and", "K", "are", "different", "in", "that", "the", "former", "is", "the", "number", "of", "topmost", "relevant", "attention", "aspects,", "whereas", "the", "latter,", "indicates", "as", "K-shot,", "refers", "to", "the", "number", "of", "samples", "in", "few-shot", "learning." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "A", "∈" ]

[ 0, 0 ]

[ "Here,", "W", "τi", "u", "indicates", "the", "selected", "k", "attention", "aspects", "of", "the", "task", "τi." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "When", "a", "novel", "task", "τi", "is", "given,", "its", "related", "attention", "aspects,", "denoted", "by", "W", "τi", "A", ",", "are", "selectively", "obtained", "by", "assigning", "the", "corresponding", "weights", "to", "members", "of", "the", "meta-level", "attention", "aspects", "WA", "in", "the", "task-adaptation", "process." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "4.2", "Top-k", "Attention", "Aspects", "Selection", "through" ]

[ 0, 0, 0, 0, 0, 0 ]

[ "In", "addition,", "AN", "and", "u", "are", "the", "total", "number", "of", "attention", "aspects", "and", "dimension", "of", "the", "attention", "aspect,", "respectively." ]

[ 0, 0, 3, 0, 3, 0, 0, 3, 4, 4, 4, 4, 0, 3, 4, 4, 4, 4, 0 ]

[ "Herein,", "we", "u", "as", "the", "meta-level", "define", "a", "matrix", "WA", "∈", "attention", "aspect", "dictionary." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "The", "meta-attention", "aspects", "in", "the", "dictionary", "are", "established", "throughout", "the", "meta-optimization", "process", "during", "which", "it", "seeks", "to", "learn", "how", "to", "attend", "according", "to", "the", "distribution", "of", "tasks." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "The", "concept", "was", "inspired", "by", "(Lin", "et", "al.,", "τi", "∼", "2017)." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "In", "this", "study,", "the", "meta-level", "knowledge", "dictionary", "maintains", "all", "attention", "aspects", "derived", "across", "tasks", "p(τ", ")." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "4.1", "Meta", "Attention", "Aspects", "Dictionary" ]

[ 0, 0, 0, 0, 0 ]

[ "The", "high-level", "operation", "is", "described", "in", "Algorithm", "1." ]

[ 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "processes:", "(1)", "deriving", "all", "valid", "meta-attention", "aspects", "across", "tasks", "(namely,", "meta-optimization),", "and", "(2)", "choosing", "a", "task-specific", "subset", "from", "all", "the", "meta-attention", "aspects", "for", "each", "task", "(called,", "task", "adaptation)." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "Output:", "WA:", "Meta-attention-aspects", "Output:" ]

[ 0, 0, 0, 0 ]

[ "Algorithm", "1", "Our", "Proposed", "Meta-Training", "tr", "Require:", "Meta", "training", "set", "Require:", "Learning-rates", "α", "(inner-update),", "β", "(outer-update)" ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3, 1, 0, 1, 0 ]

[ "It", "is", "trained", "in", "an", "end-to-end", "manner", "using", "our", "proposed", "metalearning", "strategy." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "It", "represents", "our", "meta", "learning", "framework", "for", "the", "task-specific", "feature", "extraction." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "The", "overall", "architecture", "of", "LEA", "is", "shown", "in", "Figure", "1." ]

[ 0, 0, 0, 0, 1, 0, 0, 0, 0, 0 ]

[ "4", "Proposed", "Method" ]

[ 0, 0, 0 ]

[ "Likewise,", "the", "corresponding", "[CLS]", "embedding", "is", "denoted", "as", "cτi", "j", "." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "In", "manufacturing", "a", "task-specific", "embedding,", "we", "especially", "utilize", "the", "token-level", "output", "embeddings", "of", "the", "individual", "tokens", "of", "the", "jth", "text", "instance", "under", "a", "particular", "task", "τi,", "which", "we", "denote", "as", "H", "τi", "j,L]." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "In", "this", "study,", "the", "[CLS]", "vector", "plays", "an", "important", "role", "in", "probing", "the", "distinctive", "properties", "for", "an", "incoming", "task." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "For", "downstream", "classification", "tasks,", "the", "special", "embedding", "vector", "[CLS]", "is", "typically", "used", "to", "make", "a", "prediction", "as", "the", "representative", "of", "an", "text", "instance." ]

[ 0, 0, 11, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "The", "PTMs", "end", "up", "with", "providing", "the", "corresponding", "embedding", "vectors", "(i.e.,", "denoted", "as", "[CLS]", "and", "[SEP])", "for", "the", "artificial", "tokens", "as", "well", "as", "embeddings", "for", "original", "tokens", "for", "the", "input", "text." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "Given", "a", "text", "input,", "a", "dummy", "token", "(CLS)", "is", "added", "to", "the", "beginning", "of", "the", "input,", "and", "another", "token", "(SEP)", "is", "added", "to", "the", "end", "of", "a", "sentence." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "We", "conducted", "all", "experiments", "with", "BERT", "(Devlin", "et", "al.,", "2019)", "and", "RoBERTa", "(Liu", "et", "al.,", "2019)", "as", "the", "underlying", "PTMs", "in", "the", "study." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "In", "this", "meta-testing,", "the", "dataset", "of", "task", "τi", "is", "given", "as", "Dτi", "=", "D", "3.3", "Pre-Trained", "Models" ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]