Datasets:

havens2

/

naacl2022

like 0

[ "To", "verify", "this", "guess,", "we", "test", "the", "fine-tuning", "performance", "of", "a", "random", "mapping", "with", "different", "labels", "selected." ]

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

[ "Allowing", "multiple", "labels", "can", "resolve", "mishaps", "like", "this", "and", "thus", "improve", "the", "final", "performance." ]

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

[ "This", "is", "due", "to", "the", "format", "processing", "in", "the", "pretraining", "of", "." ]

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

[ "One", "example", "is", "the", "meaningless", "</s>", "(endof-sequence", "marker)", "label", "found", "by", "AMuLaP,", "as", "shown", "in", "Table", "1." ]

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

[ "Under", "a", "few-shot", "setting,", "the", "limited", "number", "of", "training", "examples", "n", "and", "complex", "training", "procedure", "of", "the", "backbone", "model", "can", "often", "bring", "noise", "to", "both", "automatic", "label", "selection", "and", "inference." ]

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

[ "However,", "we", "find", "this", "explanation", "insufficient", "for", "understanding", "the", "mechanism", "behind", "the", "improved", "performance", "with", "multiple", "labels." ]

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

[ "Why", "Does", "AMuLaP", "Work?" ]

[ 0, 0, 1, 0 ]

[ "7", "Discussion" ]

[ 0, 0 ]

[ "With", "only", "one", "example,", "AMuLaP", "achieves", "decent", "performance", "while", "standard", "fine-tuning", "is", "close", "to", "random." ]

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

[ "In", "addition,", "the", "results", "demonstrate", "the", "effectiveness", "of", "AMuLaP", "especially", "for", "extreme", "few-shot", "settings." ]

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

[ "For", "MRPC,", "although", "the", "performance", "of", "our", "AMuLaP", "and", "standard", "fine-tuning", "fluctuate", "as", "n", "increases,", "in", "general,", "AMuLaP", "with", "fine-tuning", "can", "still", "achieve", "comparable", "performance", "to", "standard", "fine-tuning." ]

[ 0, 13, 0, 0, 0, 0, 0, 1, 0, 1, 2, 0, 0, 3, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2 ]

[ "For", "a", "harder", "task", "like", "MNLI,", "although", "the", "performance", "of", "AMuLaP", "under", "nonfinetuning", "setting", "gradually", "becomes", "saturated", "as", "n", "increases,", "AMuLaP", "under", "fine-tuning", "settings", "continues", "to", "improve", "as", "n", "increases", "and", "continues", "to", "outperform", "the", "standard", "fine-tuning." ]

[ 0, 0, 0, 0, 0, 13, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 3, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "For", "an", "easier", "task", "like", "SST-2,", "although", "only", "32", "training", "examples", "are", "used,", "the", "performance", "of", "our", "AMuLaP", "with", "non-finetuning", "and", "fine-tuning", "is", "close", "to", "saturation", "and", "can", "be", "comparable", "to", "standard", "fine-tuning", "on", "the", "entire", "dataset." ]

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

[ "When", "using", "more", "than", "16", "training", "examples,", "AMuLaP", "under", "fine-tuning", "setting", "still", "out" ]

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

[ "For", "MNLI", "and", "SST-2", "task,", "AMuLaP", "outperforms", "standard", "fine-tuning", "when", "we", "use", "no", "more", "than", "16", "training", "examples", "for", "non-finetuning", "and", "fine-tuning", "setting." ]

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

[ "Figure", "2", "illustrates", "how", "standard", "fine-tuning", "and", "our", "AMuLaP", "with", "non-finetuning", "and", "fine-tuning", "compare", "as", "n", "increases." ]

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

[ "Similarly,", "in", "this", "section,", "we", "aim", "to", "test", "how", "AMuLaP", "scales", "to", "different", "training", "set", "sizes", "n." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 3, 4, 4, 3 ]

[ "Le", "Scao", "and", "Rush", "(2021)", "explore", "the", "scaling", "law", "of", "PET", "(Schick", "and", "Schütze,", "2021a)", "when", "using", "more", "examples", "for", "training." ]

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

[ "6.3", "Scaling", "Few-Shot", "Learning" ]

[ 0, 0, 0, 0 ]

[ "An", "interesting", "exception", "is", "that", "for", "CoLA,", "the", "random", "mapping", "outperforms", "all", "label", "selection", "methods", "in", "Table", "2", "(both", "manual", "and", "automatic)", "and", "is", "close", "to", "the", "fine-tuning", "baseline." ]

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

[ "Moreover,", "a", "random", "label", "mapping", "often", "leads", "to", "lower", "performance", "than", "a", "label", "mapping", "selected", "based", "on", "the", "training", "set." ]

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

[ "Also,", "our", "multi-label", "strategy", "is", "shown", "to", "be", "effective", "at", "improving", "the", "average", "GLUE", "scores", "by", "3.6", "and", "1.1", "for", "non-finetuning", "and", "fine-tuning", "settings,", "respectively." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 13, 0, 0, 9, 0, 9, 0, 0, 0, 0, 0, 0 ]

[ "maps", "to", "two", "classes,", "optimization", "would", "be", "difficult", "due", "to", "the", "contradiction", "of", "supervision", "signals." ]

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

[ "Notably,", "deduplication", "is", "especially", "important", "for", "prompt-based", "fine-tuning", "since", "if", "the", "same", "label" ]

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

[ "For", "both", "non-finetuning", "and", "prompt-based", "fine-tuning", "settings,", "our", "deduplication", "algorithm", "can", "effectively", "improve", "the", "overall", "performance", "by", "1.1", "and", "9.9", "in", "terms", "of", "the", "GLUE", "average", "score,", "respectively." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 9, 0, 9, 0, 0, 0, 0, 13, 0, 0, 0 ]

[ "As", "shown", "in", "Table", "4,", "we", "evaluate", "the", "effect", "of", "each", "design", "choice", "on", "the", "GLUE", "benchmark." ]

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

[ "6.2", "Ablation", "Study" ]

[ 0, 0, 0 ]

[ "This", "may", "explain", "why", "AMuLaP", "works", "well,", "especially", "for", "the", "nonfinetuning", "settings." ]

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

[ "Additionally,", "we", "would", "like", "to", "highlight", "that", "AMuLaP", "mainly", "collects", "common", "words", "while", "other", "methods", "prefer", "rare", "words." ]

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

[ "Thus,", "introducing", "top-k", "truncation", "can", "resolve", "the", "problem." ]

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

[ "Although", "AMuLaP", "outputs", "three", "labels", "that", "are", "rated", "not", "suitable", "by", "the", "human", "annotator,", "it", "should", "be", "noted", "that", "all", "three", "tokens", "are", "ranked", "low", "in", "the", "candidate", "set." ]

[ 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 ]

[ "AMuLaP", "achieves", "interpretability", "that", "is", "competitive", "to", "automatic", "labels", "obtained", "by", "a", "fine-tuned", "pretrained", "language", "model,", "measured", "by", "the", "human", "agreement", "ratio." ]

[ 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 7, 8, 8 ]

[ "PETAL-CE", "suffers", "from", "incorrect", "mappings", "for", "“negative”", "while", "PETAL-LR", "occasionally", "outputs", "vague", "labels." ]

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

[ "We", "report", "the", "average", "of", "5", "runs", "along", "with", "their", "standard", "deviation", "in", "the", "parentheses." ]

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

[ "All", "of", "these", "we", "use", "are", "from", "https://github.com/princeton-nlp/", "LM-BFF/tree/main/auto_label_mapping." ]

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

[ "No", "external", "data", "used." ]

[ 0, 0, 0, 0 ]

[ "2The", "validation", "scores", "of", "all", "fine-tuned", "assignments", "are", "obtained", "on", "as", "described", "in", "Gao", "et", "al.", "(2021)." ]

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

[ "We", "shuffle", "the", "labels", "from", "each", "model", "and", "ask", "a", "human", "annotator" ]

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

[ "As", "shown", "in", "Table", "3,", "we", "list", "the", "10", "most", "likely", "label", "mappings", "output", "by", "PETAL", "(Schick", "et", "al.,", "2020),", "Auto-L", "(Gao", "et", "al.,", "2021)", "and", "AMuLaP", "for", "the", "SST-2", "dataset,", "respectively." ]

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

[ "6.1", "Case", "Study" ]

[ 0, 0, 0 ]

[ "6", "Analysis" ]

[ 0, 0 ]

[ "On", "all", "tasks", "except", "CoLA,", "AMuLaP", "outperforms", "direct", "few-shot", "fine-tuning,", "suggesting", "that", "prompting", "is", "a", "promising", "method", "for", "exploiting", "large", "pretrained", "LMs." ]

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

[ "Under", "Setting", "3,", "AMuLaP", "FT", "outperforms", "all", "baselines", "including", "Auto-L.", "Generally", "speaking,", "methods", "with", "parameter", "update", "(Setting", "3)", "have", "better", "performance", "than", "those", "that", "do", "not", "require", "access", "to", "parameters." ]

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

[ "Additionally,", "we", "attempt", "to", "replace", "the", "predicted", "token", "distribution", "of", "AMuLaP", "with", "the", "validation", "score", "of", "all", "fine-tuned", "assignments", "(Gao", "et", "al.,", "2021).2", "With", "the", "help", "of", "many", "trials", "in", "automatic", "search,", "AMuLaP", "outperforms", "Auto-L", "by", "a", "considerable", "margin", "of", "3.8", "in", "terms", "of", "the", "average", "score,", "verifying", "the", "versatility", "of", "our", "multi-label", "mechanism", "and", "label", "selection", "algorithm." ]

[ 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, 1, 0, 1, 0, 0, 0, 0, 0, 9, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "Notably,", "AMuLaP", "even", "outperforms", "Auto-L", "by", "1.3", "without", "using", "any", "external", "model", "or", "data." ]

[ 0, 1, 0, 0, 1, 0, 9, 0, 0, 0, 0, 0, 0, 0 ]

[ "Under", "Setting", "2,", "compared", "to", "variants", "of", "PETAL", "(Schick", "et", "al.,", "2020),", "AMuLaP", "has", "an", "advantage", "of", "5.8", "and", "8.5", "in", "terms", "of", "the", "average", "score", "over", "CE", "and", "LR,", "respectively." ]

[ 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 9, 0, 9, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0 ]

[ "outperforms", "GPT-3-style", "in-context", "learning", "by", "4.5", "in", "terms", "of", "the", "average", "score", "and", "outperforms", "zero-shot", "inference", "with", "manually", "designed", "labels", "by", "2.4." ]

[ 0, 1, 0, 0, 0, 9, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 9 ]

[ "Under", "Setting", "1,", "AMuLaP" ]

[ 0, 0, 0, 1 ]

[ "We", "demonstrate", "experimental", "results", "under", "three", "settings", "in", "Table", "2." ]

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

[ "5.2", "Experimental", "Results" ]

[ 0, 0, 0 ]

[ "Following", "(Gao", "et", "al.,", "2021),", "we", "grid", "search", "the", "learning", "rate", "from", "{1e-5,", "2e-5,", "5e-5}", "and", "batch", "size", "from", "{2,", "4,", "8}." ]

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

[ "For", "prompt-based", "fine-tuning", "(Setting", "3),", "where", "we", ",", "we", "search", "k", "in", "a", "smaller", "fine-tune", "the", "model", "L", "1,", "2,", "4,", "8,", "16", "due", "to", "the", "increased", "compuspace", "}", "tational", "overhead." ]

[ 0, 1, 2, 0, 0, 0, 0, 0, 0, 0, 3, 0, 0, 0, 0, 0, 0, 0, 5, 5, 5, 5, 5, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "Note", "that", "for", "settings", "that", "do", "}", "{", ",", "search", "over", "k", "is", "not", "update", "the", "parameters", "of", "fast,", "as", "we", "only", "need", "to", "run", "the", "model", "once", "and", "cache", "the", "distribution", "of", "the", "[MASK]", "token." ]

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

[ "When", "selecting", "k,", "if", "there", "are", "multiple", "k", "with", "identical", "performance", "(which", "happens", "occasionally", "given", "there", "are", "only", "16", "examples", "for", "each", "class", "in", "Ddev", "),", "we", "always", "choose", "the", "largest", "k.", "For", "Settings", "1", "and", "2,", "we", "search", "k", "over", "1,", "2,", "4,", ".", ".", "." ]

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

[ "Implementation", "Details", "We", "implement", "AMuLaP", "based", "on", "Hugging", "Face", "Transformers", "(Wolf", "et", "al.,", "2020)." ]

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

[ "This", "setting", "is", "for", "fair", "comparison", "with", "conventional", "finetuning,", "prompt-based", "fine-tuning", "with", "manual", "prompts,", "Auto-L", "(Gao", "et", "al.,", "2021)", "and", "PETAL", "(Schick", "et", "al.,", "2020)." ]

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

[ "Ddev", "in", "the", "same", "way", "as", "Setting", "2", "but", "fine-tune", "the", "parameters", "of", "the", "language", "model", "." ]

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

[ "Dtrain", "and" ]

[ 0, 0 ]

[ "•", "Setting", "3:", "We", "use" ]

[ 0, 0, 0, 0, 0 ]

[ "This", "setting", "is", "for", "fair", "comparison", "with", "Auto-L", "(Gao", "et", "al.,", "2021)", "and", "PETAL", "(Schick", "et", "al.,", "2020)." ]

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

[ "The", "paand", "an", "additional", "rameters", "of", "are", "not", "updated." ]

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

[ "•", "Setting", "2:", "We", "use", "Dtrain", "for", "label", "selection", "Ddev", "for", "k", "tuning." ]

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

[ "The", "parameters", "Ddev", "is", "not", "used." ]

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

[ "Dtrain", "alone", "for", "both", "label", "selection", "and", "tuning", "k." ]

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

[ "•", "Setting", "1:", "We", "only", "use" ]

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

[ "L" ]

[ 0 ]

[ "To", "fairly", "compare", "with", "different", "baselines,", "we", "consider", "the", "following", "three", "settings:" ]

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

[ "We", "also", "report", "the", "standard", "deviation", "for", "each", "result." ]

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

[ "Following", "Gao", "et", "al.", "(2021),", "the", "original", "development", "split", "of", "each", "dataset", "is", "used", "as", "the", "test", "set", "in", "our", "experiments." ]

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

[ "For", "each", "reported", "result,", "we", "measure", "average", "performance", "across", "5", "different", "randomly", "sampled", "Dtrain", "and", "Ddev", "splits." ]

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

[ "We", "use", "RoBERTalarge", "(Liu", "et", "al.,", "2019)", "as", "the", "backbone", "LM", "." ]

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

[ "We", "set", "k", "=", "16", "throughout", "all", "experiments." ]

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

[ "We", "sample", "n", "training", "examples", "and", "n", "development", "examples", "per", "class." ]

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

[ "Task", "Setup", "We", "closely", "follow", "the", "setup", "in", "Gao", "et", "al.", "(2021)." ]

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

[ "We", "use", "the", "officially", "released", "code", "and", "same", "hyperparameters", "for", "this", "baseline." ]

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

[ "The", "detailed", "description", "can", "be", "found", "in", "Appendix", "A.", "Note", "that", "the", "results", "of", "this", "baseline", "is", "different", "from", "those", "reported", "in", "Table", "3", "of", "Gao", "et", "al.", "(2021)", "since", "they", "search", "for", "both", "templates", "and", "label", "mapping", "whereas", "we", "fix", "the", "templates", "and", "search", "for", "the", "label", "mapping", "alone,", "for", "the", "sake", "of", "fair", "comparison." ]

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

[ "•", "Auto-L", "(Gao", "et", "al.,", "2021):", "the", "automatic", "label", "searching", "method", "with", "an", "external", "pretrained", "language", "model,", "RoBERTa-large", "(Liu", "et", "al.,", "2019)." ]

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

[ "•", "PETAL-LR", "(Schick", "et", "al.,", "2020):", "the", "variant", "of", "PETAL", "using", "the", "likelihood", "ratio", "metric." ]

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

[ "•", "PETAL-CE", "(Schick", "et", "al.,", "2020):", "the", "variant", "of", "PETAL", "using", "the", "cross-entropy", "metric." ]

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

[ "designed", "prompts", "in", "Gao", "et", "al.", "(2021)." ]

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

[ "•", "Manual", "prompts:", "we", "use", "the", "human" ]

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

[ "•", "GPT-3-style", "in-context", "learning", "(Brown", "et", "al.,", "2020):", "present", "a", "few", "examples", "to", "the", "language", "model", "and", "make", "it", "directly", "predict", "the", "next", "token", "as", "the", "prediction." ]

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

[ "•", "Majority:", "always", "predict", "the", "majority", "class" ]

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

[ "We", "compare", "our", "method", "to", "various", "baselines:" ]

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

[ "Baselines" ]

[ 0 ]

[ "The", "metrics", "for", "each", "dataset", "are", "indicated", "in", "Table", "2." ]

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

[ "We", "use", "the", "manual", "templates", "in", "Gao", "et", "al.", "(2021),", "as", "listed", "in", "Table", "1." ]

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

[ "Natural", "Language", "Inference", "Mismatched", "(MNLImm)", "(Williams", "et", "al.,", "2018),", "Question", "Natural", "Language", "Inference", "(QNLI)", "(Rajpurkar", "et", "al.,", "2016)", "and", "Recognizing", "Textual", "Entailment", "(RTE)", "(Wang", "et", "al.,", "2019)", "for", "the", "Natural", "Language", "Inference", "(NLI)", "task;", "The", "Corpus", "of", "Linguistic", "Acceptability", "(CoLA)", "(Warstadt", "et", "al.,", "2019)", "for", "Linguistic", "Acceptability." ]

[ 0, 0, 0, 0, 13, 0, 0, 0, 0, 13, 14, 14, 14, 13, 0, 0, 0, 0, 0, 13, 14, 14, 13, 0, 0, 0, 0, 0, 0, 11, 12, 12, 11, 0, 13, 14, 14, 14, 14, 13, 0, 0, 0, 0, 0, 11, 12 ]

[ "We", "report", "the", "average", "of", "5", "runs", "along", "with", "their", "standard", "deviation", "in", "the", "parentheses." ]

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

[ "Specifically,", "we", "test", "on", "Microsoft", "Research", "Paraphrase", "Matching", "(MRPC)", "(Dolan", "and", "Brockett,", "2005),", "Quora", "Question", "Pairs", "(QQP)", "for", "Paraphrase", "Similarity", "Matching;", "Stanford", "Sentiment", "Treebank", "(SST-2)", "(Socher", "et", "al.,", "2013)", "for", "Sentiment", "Classification;", "Multi-Genre", "Natural", "Language", "Inference", "Matched", "(MNLI-m),", "Multi-Genre" ]

[ 0, 0, 0, 0, 13, 14, 14, 14, 13, 0, 0, 0, 0, 13, 14, 14, 13, 0, 11, 12, 12, 13, 14, 14, 13, 0, 0, 0, 0, 0, 11, 12, 13, 14, 14, 14, 14, 13, 0 ]

[ "Datasets", "We", "evaluate", "seven", "classification", "tasks", "of", "the", "GLUE", "benchmark", "(Wang", "et", "al.,", "2019)." ]

[ 0, 0, 0, 0, 11, 0, 0, 0, 13, 0, 0, 0, 0, 0 ]

[ "5.1", "Experimental", "Setting" ]

[ 0, 0, 0 ]

[ "5", "Experiments" ]

[ 0, 0 ]

[ "Then,", "we", "can", "truncate", "the", "candidate", "sets", "of", "each", "class", "and", "select", "the", "k", "most", "likely", "tokens", "from", "each", "set." ]

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

[ "To", "ensure", "this,", "we", "have", "to", "iterate", "over", "the", "vocabulary", "and", "check", "that", "for", "every", "token." ]

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

[ "Thus,", "we", "want", "to", "make", "sure", "that", "(2)", "Each", "token", "only", "belongs", "to", "at", "most", "one", "label", "set", "where", "it", "has", "the", "highest", "probability." ]

[ 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", "example,", "if", "we", "simply", "take", "the", "10", "most", "likely", "tokens", "for", "the", "SST-2", "dataset", "(Socher", "et", "al.,", "2013),", "we", "would", "find", "“good”", "in", "both", "positive", "and", "negative", "label", "sets,", "although", "it", "is", "ranked", "second", "place", "in", "the", "positive", "set", "and", "ninth", "in", "the", "negative", "set." ]

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

[ "However,", "in", "practice,", "we", "would", "find", "common", "words", "in", "more", "than", "one", "label", "set." ]

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

[ "A", "simple", "solution", "is", "to", "select", "the", "k", "most", "likely", "tokens", "predicted", "for", "the", "[MASK]", "token", "in", "the", "training", "examples", "of", "each", "class", "y." ]

[ 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", "example,", "in", "a", "sentiment", "classification", "task,", "we", "would", "like", "to", "see", "positive", "words", "in", "the", "label", "set", "of", "the", "“positive”", "class", "while", "negative", "words", "in", "the", "label", "set", "of", "the", "“negative”", "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, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]