Datasets:

havens2

/

naacl2022

like 0

[ "We", "aim", "to", "achieve", "two", "goals:", "(1)", "Selecting", "the", "most", "likely", "label", "mapping", "based", "on", "the", "training", "set." ]

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

[ "Finally,", "for", "each", "class,", "we", "choose", "the", "top-k", "tokens", "as", "label", "words." ]

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

[ "Then", "each", "token", "in", "is", "assigned", "to", "the", "class", "with", "the", "highest", "probability", "(e.g.,", "the", "token", "terrible", "is", "assigned", "to", "the", "class", "negative,", "the", "token", "great", "is", "assigned", "to", "the", "class", "positive)." ]

[ 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 ]

[ "All", "the", "obtained", "probability", "distributions", "are", "summed", "by", "class", "and", "probability", "distribution", "over", "the", "vocabulary", "normalized", "to", "get", "the", "probability", "distribution", "of", "each", "class." ]

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

[ "Each", "training", "sample", "with", "the", "task-specific", "template", "(the", "underlined", "text)", "is", "fed", "into", "a", "pretrained", "language", "model", "to", "get", "its", "own", "." ]

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

[ "5486" ]

[ 0 ]

[ "We", "them", "and", "selecting", "the", "best", "one", "with", "introduce", "a", "new", "selection", "algorithm", "for", "label", "mapping", "that", "achieves", "competitive", "results", "compared", "to", "previous", "efforts." ]

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

[ "(Gao", "et", "al.,", "2021)", "exploits", "a", "large", "pretrained", "masked", "language", "model", "(RoBERTa,", "Liu", "et", "al.,", "2019)", "to", "construct", "a", "pruned", "set", "of", "label", "words", "and", "then", "determine", "the", "final", "mapping", "by", "fine-tuning", "on", "all", "of", "Ddev", "." ]

[ 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 ]

[ "(2020)", "explores", "automatic", "label", "mapping", "searching", "but", "it", "still", "requires", "manual", "pre-filtering", "and", "significantly", "underperforms", "the", "manual", "mapping." ]

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

[ "Previously,", "Schick", "and", "Schütze", "(2021a,b)", "both", "use", "hand-crafted", "label", "mappings", "while", "Schick", "et", "al." ]

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

[ "Selecting", "a", "good", "label", "mapping", "often", "requires", "significant", "human", "effort,", "including", "domain", "knowledge", "and", "trial-and-error." ]

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

[ "Finding", "a", "good", "label", "mapping", "is", "non-trivial,", "esM", "pecially", "when", "M′", "maps", "an", "original", "label", "to", "a", "set", "of", "label", "words", "instead", "of", "one." ]

[ 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", "Automatic", "Label", "Selection" ]

[ 0, 0, 0, 0 ]

[ "l", "=" ]

[ 0, 0 ]

[ "Then,", "we", "can", "simply", "make", "predictions", "by", "selecting", "the", "label", "with", "the", "largest", "likelihood." ]

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

[ "We", "denote", "the", "k.", "For", "class", ",", "the", "predicted", "probability", "is", "calculated", "as:" ]

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

[ "Instead", "of", "a", "one-to-one", "mapping", "from", "the", "origto", "to", "(y)", "of", "k", "words." ]

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

[ "This", "design", "has", "a", "similar", "advantage", "of", "exploiting", "multiple", "labels", "without", "training", "and", "ensembling", "multiple", "models." ]

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

[ "We", "instead", "use", "a", "straightforward", "sum", "to", "consider", "multiple", "label", "words", "when", "making", "predictions." ]

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

[ "Schick", "et", "al.", "(2020)", "use", "multiple", "label", "combinations", "for", "PET", "(Schick", "and", "Schütze,", "2021a)", "and", "ensemble", "them", "afterwards." ]

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

[ "This", "can", "be", "resolved", "by", "introducing", "multiple", "label", "words." ]

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

[ "We", "also", "argue", "that", "selecting", "only", "one", "label", "(especially", "automatically)", "may", "bring", "noise." ]

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

[ "Selecting", "one", "label", "word", "can", "be", "insufficient", "for", "some", "complicated", "tasks,", "as", "mentioned", "in", "Schick", "et", "al.", "(2020)." ]

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

[ "4.1", "Exploiting", "Multiple", "Labels" ]

[ 0, 0, 0, 0 ]

[ "4", "Automatic", "Multi-Label", "Prompting" ]

[ 0, 0, 0, 0 ]

[ "|", "Alternately,", "one", "can", "further", "fine-tune", "pervised", "pairs", "performance." ]

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

[ "The", "templates", "used", "for", "prompting", "are", "from", "Gao", "et", "al.", "(2021)." ]

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

[ "and" ]

[ 0 ]

[ "Since", "complete", "the", "[MASK]", "token", "in", "an", "input", "sequence,", "we", "can", "directly", "make", "zero-shot", "prediction", "of", "the", "probability", "of", "class", "y", "by", "the", "masked", "language" ]

[ 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 ]

[ "Note", "that", "since", "we", "T", "focus", "on", "automatically", "finding", "the", "label", "mapping", "from", "Gao", "et", "al.", ",", "we", "use", "the", "manual", "templates", "M", "T", "is", "trained", "to", "(2021)", "throughout", "this", "paper." ]

[ 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", "then", "map", "the", "original", "label", "space", "to", "a", "set", "of", "selected", "words", "from", "the", "vocabuY", "→", "V", "′.", "Some", "examples", "of", "lary,", "denoted", "as", "are", "shown", "in", "Table", "1." ]

[ 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 ]

[ "For", "an", "input", "example", "x", "(a", "single", "sentence", "or", "a", "sentence", "pair),", "we", "first", "use", "a", "task-specific", "temto", "convert", "it", "to", "x′,", "a", "token", "sequence", "with", "plate", "a", "[MASK]", "token." ]

[ 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 ]

[ "Thus,", "we", "use", "a", "few-shot", "development", "set", "with", "the", "same", "size", "as", "the", "training", "set", "(i.e.,", "),", "to", "be", "consistent", "with", "Gao", "et", "al.", "(2021)", "and", "constitute", "a", "“true", "few-shot”", "setting", "(Perez", "et", "al.,", "2021)." ]

[ 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 ]

[ "As", "pointed", "out", "in", "Perez", "et", "al.", "(2021),", "using", "the", "full", "development", "set", "may", "be", "misleading", "to", "claim", "a", "few-shot", "setting." ]

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

[ "Given", "a", "pretrained", "and", "its", "defined", "label", "language", "model", ",", "we", "have", "n", "training", "examples", "per", "class", "space", "for", "the", "training", "set", "Dtrain", "." ]

[ 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", "follow", "the", "setup", "in", "LM-BFF", "(Gao", "et", "al.,", "2021)", "for", "few-shot", "text", "classification." ]

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

[ "3", "Prompting", "for", "Few-Shot", "Classification" ]

[ 0, 0, 0, 11, 12 ]

[ "Different", "from", "these", "works,", "AMuLaP", "is", "a", "discrete", "prompting", "method", "that", "has", "better", "interpretability", "and", "works", "well", "in", "the", "few-shot", "setting." ]

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

[ "Notably,", "different", "from", "discrete", "prompting,", "these", "works", "often", "use", "all", "training", "data", "to", "update", "model", "weights." ]

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

[ "Guo", "et", "al.", "(2021)", "introduces", "Q-Learning", "to", "optimize", "the", "soft", "prompt." ]

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

[ "Lester", "et", "al.", "(2021)", "reveals", "that", "prompt", "tuning", "is", "more", "competitive", "when", "scaled", "up", "and", "can", "achieve", "identical", "performance", "to", "conventional", "fine-tuning", "when", "the", "model", "is", "large", "enough." ]

[ 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 ]

[ "Li", "and", "Liang", "(2021)", "applies", "a", "similar", "method", "for", "natural", "language", "generation", "and", "achieves", "comparable", "performance", "to", "fine-tuning", "while", "updating", "only", "0.1%", "of", "model", "parameters." ]

[ 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 ]

[ "Zhong", "et", "al.", "(2021)", "and", "Qin", "and", "Eisner", "(2021)", "propose", "continuous", "prompts", "for", "knowledge", "probing", "by", "tuning", "some", "trainable", "vectors", "in", "the", "input", "sequence", "while", "fixing", "the", "rest", "of", "the", "input." ]

[ 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", "parallel", "with", "text-based", "Continuous", "Prompts", "discrete", "prompts,", "there", "is", "also", "a", "line", "of", "work", "focused", "on", "tuning", "only", "a", "fraction", "of", "parameters", "of", "an", "LM", "with", "the", "help", "of", "continuous", "prompts", "(i.e.,", "soft", "prompts)." ]

[ 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 ]

[ "by", "manually", "writing", "prompt", "templates", "and", "shows", "that", "a", "large", "language", "model", "with", "multitask", "training", "can", "generalize", "to", "unseen", "tasks." ]

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

[ "T0", "(Sanh", "et", "al.,", "2022;", "Bach", "et", "al.,", "2022)", "constructs", "a", "dataset", "of", "different", "NLP", "tasks" ]

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

[ "Concurrently", "to", "our", "work,", "Hu", "et", "al.", "(2021)", "propose", "a", "method", "that", "exploits", "an", "external", "knowledge", "base", "to", "find", "label", "mapping." ]

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

[ "Different", "from", "these", "prior", "studies,", "our", "proposed", "AMuLaP", "is", "a", "simple", "and", "interpretable", "method", "for", "few-shot", "prompting", "that", "can", "work", "well", "with", "and", "without", "access", "to", "model", "weights." ]

[ 0, 0, 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 ]

[ "These", "searched", "prompts", "and", "labels", "are", "often", "uninterpretable", "by", "humans." ]

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

[ "AutoPrompt", "uses", "discretization", "techniques", "to", "approximately", "map", "a", "continuous", "vector", "back", "to", "tokens", "in", "the", "vocabulary", "(i.e.,", "“vocablization”)." ]

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

[ "PET", "and", "its", "variants", "also", "require", "a", "large", "unlabeled", "set", "and", "need", "to", "be", "fine-tuned", "multiple", "times." ]

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

[ "However,", "these", "methods", "require", "parameter", "updates", "with", "gradient", "descent,", "which", "is", "infeasible", "without", "access", "to", "the", "model", "weights", "(e.g.,", "GPT-3)." ]

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

[ "PETAL", "(Schick", "et", "al.,", "2020)", "augments", "Pattern", "Exploiting", "Training", "(PET,", "Schick", "and", "Schütze,", "2021a,b)", "with", "automatically", "identified", "label", "words;", "Gao", "et", "al.", "(2021)", "uses", "re-ranking", "to", "find", "the", "best", "label", "words", "by", "fine-tuning", "a", "RoBERTa", "model", "on", "the", "candidates", "searched", "by", "RoBERTa,", "and", "using", "an", "external", "generation", "model", "for", "data", "augmentation", "of", "prompt", "templates;", "AutoPrompt", "(Shin", "et", "al.,", "2020)", "uses", "a", "gradient-based", "search", "to", "determine", "both", "prompts", "and", "label", "words." ]

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

[ "Le", "Scao", "and", "Rush", "(2021)", "analyzes", "the", "effect", "of", "prompting", "and", "concludes", "that", "a", "single", "prompt", "may", "be", "worth", "100", "training", "examples", "in", "fine-tuning." ]

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

[ "Recent", "works", "(Schick", "and", "Schütze,", "2021a,b;", "Gao", "et", "al.,", "2021)", "demonstrate", "that", "even", "smaller", "PLMs", "have", "similar", "few-shot", "learning", "capacity." ]

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

[ "Recent", "works", "(Petroni", "et", "al.,", "2019;", "Davison", "et", "al.,", "2019;", "Jiang", "et", "al.,", "2020)", "design", "prompts", "to", "probe", "the", "factual", "and", "commonsense", "knowledge", "encoded", "within", "a", "PLM." ]

[ 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 ]

[ "Brown", "et", "al.", "(2020)", "proposes", "an", "intuitive", "in-context", "learning", "paradigm", "by", "concatenating", "a", "few", "input", "and", "output", "examples", "and", "feeding", "them", "to", "the", "language", "model", "and", "let", "the", "model", "autoregressively", "generate", "answers", "for", "new", "examples." ]

[ 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 ]

[ "The", "release", "of", "GPT-3", "(Brown", "et", "al.,", "2020)", "has", "led", "to", "interest", "in", "prompting,", "a", "new", "way", "to", "leverage", "pretrained", "language", "models", "(PLM)." ]

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

[ "Discrete", "Prompts" ]

[ 0, 0 ]

[ "2", "Related", "Work" ]

[ 0, 0, 0 ]

[ "why", "does", "AMuLaP", "work", "and", "discuss", "the", "pros", "and", "cons", "of", "prompting", "as", "a", "new", "paradigm." ]

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

[ "Moreover,", "we", "attempt", "to", "scale", "AMuLaP", "with", "different", "sizes", "of", "the", "training", "set", "and", "find", "AMuLaP", "to", "work", "surprisingly", "well", "even", "with", "one", "or", "two", "shots." ]

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

[ "We", "conduct", "extensive", "experiments", "and", "demonstrate", "the", "effectiveness", "of", "our", "method", "under", "multiple", "settings." ]

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

[ "Compared", "with", "a", "hand-crafted", "label", "mapping", "and", "previous", "works", "on", "automatic", "label", "mapping", "(Schick", "et", "al.,", "2020;", "Gao", "et", "al.,", "2021),", "AMuLaP", "achieves", "competitive", "performance", "despite", "being", "simpler", "and", "does", "not", "require", "access", "to", "the", "weights", "of", "the", "backbone", "model,", "or", "finetune", "an", "external", "pretrained", "language", "model", "for", "searching", "label", "mapping." ]

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

[ "AMuLaP", "exploits", "multiple", "labels", "to", "suppress", "the", "noise", "and", "inherently", "extend", "the", "training", "set", "for", "prompt-based", "fine-tuning." ]

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

[ "AMuLaP", "is", "a", "parameter-free", "statistical", "technique", "that", "can", "identify", "the", "label", "patterns", "from", "a", "few-shot", "training", "set", "given", "a", "prompt", "template." ]

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

[ "We", "propose", "Automatic", "Multi-Label", "Prompting", "(AMuLaP),", "a", "simple", "yet", "effective", "method", "to", "tackle", "the", "label", "selection", "problem", "for", "few-shot", "classification." ]

[ 0, 0, 1, 2, 2, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 11, 12 ]

[ "In", "this", "paper,", "we", "aim", "to", "design", "a", "method", "that", "can", "automatically", "find", "a", "good", "label", "mapping", "to", "save", "human", "effort", "from", "label", "engineering." ]

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

[ "Thus,", "an", "efficient", "automatic", "label", "mapping", "method", "is", "desirable." ]

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

[ "One", "may", "argue", "that", "the", "same", "effort", "can", "be", "used", "to", "label", "more", "supervised", "data", "for", "a", "conventional", "deep", "learning", "pipeline." ]

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

[ "However,", "manually", "assigning", "the", "label", "mapping", "requires", "human", "expertise", "with", "trial", "and", "error." ]

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

[ "Meanwhile,", "finding", "a", "good", "mapping", "from", "the", "original", "task", "labels", "to", "tokens", "(i.e.,", "label", "engineering)", "is", "also", "critical", "to", "few-shot", "performance,", "as", "found", "in", "Schick", "et", "al.", "(2020);", "Gao", "et", "al.", "(2021)." ]

[ 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 ]

[ "Many", "attempts", "have", "been", "made", "in", "this", "emerging", "direction", "of", "prompt", "engineering", "(Shin", "et", "al.,", "2020;", "Gao", "et", "al.,", "2021)." ]

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

[ "In", "prompting,", "the", "design", "of", "prompts", "often", "plays", "an", "important", "role." ]

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

[ "Prompt-based", "learning", "(i.e.,", "prompting)", "aims", "to", "use", "a", "template", "to", "convert", "the", "original", "input", "into", "a", "prompt-based", "input", "with", "some", "unfilled", "masked" ]

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

[ "Furthermore,", "recent", "works", "(Schick", "and", "Schütze,", "2021a,b;", "Gao", "et", "al.,", "2021)", "show", "that", "prompts", "can", "also", "help", "the", "model", "generalize", "better", "in", "fine-tuning." ]

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

[ "The", "use", "of", "prompts", "can", "strengthen", "the", "explicit", "connection", "between", "input", "and", "output,", "helping", "the", "model", "exploit", "the", "knowledge", "learned", "from", "pretraining", "in", "a", "better", "way." ]

[ 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", "setting", "of", "few-shot", "learning", "is", "closer", "to", "how", "humans", "learn", "to", "solve", "a", "task,", "often", "without", "many", "examples", "as", "in", "a", "traditional", "deep", "learning", "paradigm." ]

[ 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 ]

[ "These", "works", "demonstrate", "the", "potential", "of", "using", "natural", "language", "prompts", "to", "encourage", "the", "model", "to", "recall", "similar", "patterns", "in", "its", "training", "corpus", "and", "thus", "make", "accurate", "predictions." ]

[ 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 ]

[ "Since", "the", "release", "of", "GPT-3", "(Brown", "et", "al.,", "2020),", "several", "studies", "have", "focused", "on", "exploiting", "pretrained", "language", "models", "with", "only", "a", "few", "training", "examples", "(Brown", "et", "al.,", "2020;", "Gao", "et", "al.,", "2021;", "Shin", "et", "al.,", "2020)." ]

[ 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 ]

[ "Introduction" ]

[ 0 ]

[ "Our", "experiments", "demonstrate", "that", "AMuLaP", "achieves", "competitive", "performance", "on", "the", "GLUE", "benchmark", "without", "human", "effort", "or", "external", "resources.1" ]

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

[ "Our", "method", "exploits", "one-to-many", "label", "mappings", "and", "a", "statistics-based", "algorithm", "to", "select", "label", "mappings", "given", "a", "prompt", "template." ]

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

[ "In", "this", "paper,", "we", "propose", "Automatic", "Multi-Label", "Prompting", "(AMuLaP),", "a", "simple", "yet", "effective", "method", "to", "automatically", "select", "label", "mappings", "for", "few-shot", "text", "classification", "with", "prompting." ]

[ 0, 0, 0, 0, 0, 1, 2, 2, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 11, 12, 12, 0, 0 ]

[ "Prompt-based", "learning", "(i.e.,", "prompting)", "is", "an", "emerging", "paradigm", "for", "exploiting", "knowledge", "learned", "by", "a", "pretrained", "language", "model." ]

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

[ "Abstract" ]

[ 0 ]

[ "Simple", "and", "Interpretable", "Few-Shot", "Classification" ]

[ 0, 0, 0, 11, 12 ]

[ "Automatic", "Multi-Label", "Prompting:" ]

[ 1, 2, 2 ]

[ "One", "should", "be", "careful", "to", "apply", "our", "proposed", "model", "and", "all", "other", "prompt", "tuning", "methods", "in", "high-stakes", "areas", "without", "a", "comprehensive", "test." ]

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

[ "However,", "if", "the", "backbone", "model", "stored", "online", "is", "attacked,", "whether", "IDPG", "could", "still", "work", "well", "remains", "unknown." ]

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

[ "It", "tunes", "small", "portion", "parameters", "while", "directly", "employs", "backbone", "model", "parameters", "without", "any", "changing." ]

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

[ "Our", "proposed", "model", "IDPG", "is", "a", "novel", "efficient", "transfer", "learning", "method." ]

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

[ "A.8", "Potential", "Risks" ]

[ 0, 0, 0 ]

[ "IDPG", "shows", "its", "stability", "when", "scaling", "to", "larger", "models", "with", "longer", "prompts." ]

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

[ "We", "present", "the", "impact", "of", "prompt", "length", "among", "several", "prompt", "tuning", "methods", "in", "Figure", "9." ]

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

[ "A.7", "Ablation", "Study", "on", "Prompt", "Length" ]

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

[ "We", "present", "the", "cosine", "similarity", "distributions", "when", "k", "=", "100", "and", "k", "=", "300", "in", "Figure", "8a", "and", "in", "Figure", "8b,", "respectively." ]

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

[ "A.6", "Cosine", "Similarity", "Distributions", "in", "STS-B" ]

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

[ "It", "is", "interesting", "that", "IDPG", "achieves", "best", "results", "in", "position", "0", "while", "the", "standard", "prompt-tuning", "achieves", "the", "best", "results", "in", "position", "4", "for", "both", "single-sentence", "and", "sentence-pair", "tasks." ]

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

[ "Figure", "7(b)", "illustrates", "the", "effect", "of", "prompt", "position", "on", "the", "supplementary", "training", "phase." ]

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