Datasets:

havens2

/

naacl2022

like 0

[ "We", "come", "to", "a", "similar", "conclusion", "that", "multi-layer", "instancedependent", "prompt", "tuning", "model", "(M-IDPG)", "is", "significantly", "better", "than", "the", "single-layer", "method", "(SIDPG)", "in", "both", "evaluation", "settings." ]

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

[ "To", "boost", "single-layer", "IDPG", "performance,", "we", "add", "supplementary", "training", "(cf.", "Appendix", "A.4)", "and", "conduct", "ablation", "studies", "in", "Appendix", "A.5." ]

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

[ "P-tuning", "v2", "(Liu", "et", "al.,", "2021a)", "conducted", "substantial", "ablation", "studies", "on", "the", "influence", "of", "inserting", "prompt", "into", "different", "transformer", "layers." ]

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

[ "Multi-layer?" ]

[ 0 ]

[ "4.5.4", "Prompt", "Insertion:", "Single-layer", "or" ]

[ 0, 0, 0, 0, 0 ]

[ "Detailed", "information", "for", "all", "models’", "performance", "on", "each", "task", "can", "be", "found", "in", "Appendix", "A.3." ]

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

[ "In", "Table", "1,", "M-IDPG-PHM", "uses", "the", "previous", "layer’s", "output", "as", "input,", "M", "version", "as", "the", "generator,", "and", "16", "as", "the", "generator", "hidden", "size." ]

[ 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 5, 0, 0, 0, 3, 4 ]

[ "As", "for", "the", "generator", "selection,", "the", "three", "models", "perform", "as", "expected", "(S", "version", "<", "M", "version", "<", "L", "version)." ]

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

[ "However,", "as", "shown", "in", "Figure", "3,", "the", "experiment", "results", "suggest", "no", "significant", "difference", "between", "the", "two", "input", "ways." ]

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

[ "In", "a", "multi-layer", "case,", "the", "input", "to", "each", "layer", "generator", "has", "another", "option,", "i.e.,", "the", "previous", "layer’s", "output." ]

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

[ "in", "single-layer", "prompt", "generation", "model,", "the", "input", "to", "G", "is", "M(xi)", "-", "the", "representation", "of", "input", "sequence", "xi." ]

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

[ "We", "hypothesize", "that", "the", "smaller", "hidden", "size", "of", "16", "is", "already", "enough", "to", "store", "useful", "instance", "information,", "and", "setting", "m", "too", "large", "may", "be", "less", "efficient." ]

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

[ "Surprisingly,", "we", "find", "that", "generator", "with", "a", "hidden", "size", "16", "is", "not", "far", "from", "the", "large", "model", "(92.0", "vs.", "92.1,", "respectively,", "in", "M", "version)." ]

[ 0, 0, 0, 0, 0, 0, 0, 3, 4, 5, 0, 0, 0, 0, 0, 0, 0, 9, 0, 9, 0, 0, 0, 0 ]

[ "We", "compare", "two", "models", "with", "m", "=", "16", "and", "m", "=", "256." ]

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

[ "Another", "way", "to", "reduce", "the", "training", "parameters", "is", "by", "adjusting", "the", "hidden", "size", "m", "of", "the", "generator." ]

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

[ "If", "each", "transformer", "layer", "requires", "an", "independent", "generator", "Gi,", "the", "number", "of", "training" ]

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

[ "When", "applying", "the", "instance-dependent", "generation", "model", "G", "into", "a", "multi-layer", "case,", "the", "first", "challenge", "we", "face", "is", "the", "considerable", "increase", "in", "training", "parameters." ]

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

[ "4.5.3", "Multi-layer", "Architecture", "Exploration" ]

[ 0, 0, 0, 0 ]

[ "On", "the", "other", "hand,", "this", "ablation", "study", "further", "verifies", "PHM", "layers’", "efficiency", "in", "the", "generation", "model." ]

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

[ "We", "observe", "that", "including", "DNN", "as", "a", "generator", "doesn’t", "improve", "performance", "signif" ]

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

[ "Hence,", "we", "compare", "the", "PHM-based", "prompt", "generator", "with", "the", "DNN-based", "prompt", "generator,", "as", "shown", "in", "Table", "1." ]

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

[ "An", "open", "question", "we", "seek", "to", "answer", "is", "what", "is", "the", "best", "generation", "model", "for", "prompt", "regardless", "of", "training", "parameters." ]

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

[ "To", "reduce", "the", "tuning", "parameters,", "we", "substitute", "the", "DNN", "layers", "with", "PHM", "layers." ]

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

[ "4.5.2", "Prompt", "Generator:", "PHM", "or", "DNN?" ]

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

[ "Adopting", "GloVe", "as", "sentence", "encoder", "would", "avoid", "going", "through", "the", "LM", "twice,", "thus", "effectively", "reducing", "IDPG’s", "run-time", "complexity", "by", "half." ]

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

[ "One", "of", "the", "drawbacks", "of", "our", "method", "is", "that", "it", "is", "twice", "as", "expensive", "to", "run", "compared", "to", "Compacter,", "even", "though", "it", "uses", "slightly", "fewer", "parameters." ]

[ 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 ]

[ "According", "to", "Table", "1,", "using", "GloVe", "as", "sentence", "encoder", "to", "generate", "prompts", "doesn’t", "sacrifice", "much", "performance", "over", "the", "ten", "tasks", "and", "outperforms", "prompt", "tuning", "and", "P-tuning", "v2." ]

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

[ "Method" ]

[ 0 ]

[ "For", "all", "other", "tasks,", "we", "report", "accuracy." ]

[ 0, 0, 0, 0, 0, 0, 7 ]

[ "We", "report", "the", "average", "of", "accuracy", "and", "F1", "for", "both", "MRPC", "and", "QQP,", "and", "average", "of", "Pearson", "and", "Spearman", "correlation", "coefficients", "for", "STS-B." ]

[ 0, 0, 0, 0, 0, 7, 0, 7, 0, 0, 13, 0, 13, 0, 0, 0, 7, 8, 8, 8, 8, 0, 13 ]

[ "Underline", "marks", "the", "best", "result", "among", "all", "prompt", "tuning", "methods." ]

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

[ "Bold", "marks", "the", "best", "result", "among", "all", "competing", "methods." ]

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

[ "We", "report", "average", "results", "across", "5", "runs", "with", "different", "initialization." ]

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

[ "Specifically,", "we", "take", "the", "average", "of", "word", "vectors", "as", "the", "sentence", "embeddings:" ]

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

[ "To", "answer", "this", "question,", "we", "apply", "the", "pre-trained", "GloVe", "word", "vectors2", "to", "extract", "the", "sentence", "representation." ]

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

[ "One", "open", "question", "is", "to", "explore", "reliability", "on", "lightweight", "sentence", "representations", "such", "as", "GloVe", "embedding", "(Pennington", "et", "al.,", "2014)", "or", "token", "embedding", "of", "pre-trained", "language", "models." ]

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

[ "Obtaining", "contextualized", "transformer", "sentence", "embedding", "is", "often", "expensive", "if", "it", "is", "not", "pre-computed." ]

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

[ "The", "proposed", "IDPG", "method", "relies", "on", "pre-trained", "LM", "to", "extract", "sentence", "representation,", "i.e.,", "[CLS]", "token", "embedding." ]

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

[ "4.5.1", "Sentence", "Encoder:", "GloVe", "or", "LMs?" ]

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

[ "We", "conduct", "several", "ablation", "studies", "including", "exploration", "of", "different", "generator", "architectures", "and", "impact", "of", "selecting", "different", "prompt", "positions." ]

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

[ "4.5", "Intrinsic", "Study" ]

[ 0, 0, 0 ]

[ "Testing", "the", "limitation", "of", "our", "model", "without", "freezing", "any", "parameters", "would", "be", "an", "interesting", "investigation,", "but", "is", "not", "the", "main", "focus", "of", "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 ]

[ "We", "want", "to", "highlight", "that", "we", "are", "exploring", "a", "solution", "by", "training", "as", "few", "parameters", "as", "possible", "while", "maintaining", "good", "performance." ]

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

[ "We", "also", "note", "that", "other", "state-of-the-art", "models,", "such", "as", "LM-BFF", "(Gao", "et", "al.,", "2021a),", "attempt", "to", "address", "the", "few-shot", "learning", "problem", "from", "a", "different", "perspective." ]

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

[ "We", "suspect", "that", "this", "may", "be", "due", "to", "poor", "initialization", "leading", "the", "model", "to", "non-optimal", "parameters." ]

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

[ "We", "also", "observe", "that", "sometimes", "when", "K", "is", "small,", "our", "method", "results", "have", "high", "variance", "(e.g.,", "4.6", "on", "MPQA,", "when", "K", "=", "100)." ]

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

[ "When", "K", "becomes", "larger,", "IDPG-PHM", "still", "maintains", "good", "results", "with", "1.9pt", "and", "0.2pt", "improvement", "(K=500);", "and", "2.0pt", "and", "0.2pt", "improvement", "(K=1000)", "in", "accuracy", "with", "traditional", "prompt", "tuning", "and", "P-tuning", "v2", "approaches,", "respectively." ]

[ 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 9, 0, 9, 0, 0, 0, 9, 0, 9, 0, 0, 0, 7, 0, 1, 2, 2, 0, 1, 2, 0, 0 ]

[ "This", "improvement", "illustrates", "that", "our", "method", "has", "better", "generalization", "in", "few-shot", "settings." ]

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

[ "In", "the", "extreme", "low-resource", "case", "when", "K=100,", "M-IDPG-PHM", "performs", "2.5pt", "better", "than", "the", "traditional", "prompt", "tuning", "method", "and", "0.5pt", "better", "than", "the", "multi-layer", "P-Tuning", "v2", "method." ]

[ 0, 0, 0, 0, 0, 0, 0, 1, 0, 9, 0, 0, 0, 0, 0, 0, 0, 0, 9, 0, 0, 0, 0, 1, 2, 0 ]

[ "Following", "the", "existing", "evaluation", "protocols", "in", "the", "few-shot", "setting", "(He", "et", "al.,", "2021),", "we", "sample", "a", "subset", "of", "the", "training", "data", "for", "each", "task", "with", "size", "K", "as", "our", "∈", "{", "training", "data", "and", "another", "subset", "with", "size", "1000", "as", "a", "development", "set." ]

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

[ "4.4", "Performance", "in", "low-resource", "scenario" ]

[ 0, 0, 0, 0, 0 ]

[ "#", "Parameters" ]

[ 0, 0 ]

[ "Method" ]

[ 0 ]

[ "Our", "proposed", "method,", "especially", "the", "M-IDPGPHM,", "falls", "in", "the", "gap", "between", "prompt-tuning", "and", "adapter", "model,", "since", "it", "only", "requires", "training", "134K", "parameters", "and", "performs", "on", "par", "with", "Compacter." ]

[ 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 ]

[ "However,", "its", "performance", "is", "worse", "than", "a", "lightweight", "adapter", "model", "(e.g.,", "Compacter", "with", "149K", "parameters)." ]

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

[ "Traditional", "prompt-tuning", "method", "only", "requires", "training", "a", "token", "embedding", "table", "with", "a", "few", "thousand", "parameters." ]

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

[ "The", "general", "goal", "for", "efficient", "transfer", "learning", "is", "to", "train", "models", "with", "fewer", "parameters", "while", "achieving", "better", "performance." ]

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

[ "Table", "2", "lists", "the", "number", "of", "trainable", "parameters", "for", "different", "methods", "excluding", "the", "classification", "head." ]

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

[ "4.3", "Efficiency" ]

[ 0, 0 ]

[ "(vi)", "When", "we", "fix", "the", "training", "parameters", "to", "be", "the", "same,", "the", "comparison", "between", "Prompttuning-134", "and", "M-IDPG-PHM", "illustrates", "that", "our", "approach", "works", "better", "than", "prompt", "tuning", "not", "just", "because", "of", "using", "more", "parameters." ]

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

[ "(v)", "GloVe-based", "sentence", "encoder", "also", "performs", "similar", "to", "LMbased", "sentence", "encoder,", "indicating", "the", "advancement", "of", "instance-dependent", "prompt", "generation", "does", "not", "rely", "on", "a", "robust", "contextual", "sentence", "encoder." ]

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

[ "based", "generator", "performs", "on", "par", "with", "the", "DNNbased", "generator", "while", "having", "a", "significantly", "lower", "number", "of", "trainable", "parameters." ]

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

[ "Specifically,", "M-IDPG-PHM", "performs", "0.84pt", "and", "0.36pt", "better", "than", "RoBERTa", "and", "EFL,", "respectively." ]

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

[ "The", "four", "best", "results", "(MPQA,", "Subj,", "CR,", "MR)", "among", "all", "competing", "methods", "in", "single-sentence", "classification", "tasks", "are", "made", "by", "IDPG", "models." ]

[ 0, 0, 0, 0, 13, 13, 13, 13, 0, 0, 0, 0, 0, 11, 12, 0, 0, 0, 0, 1, 0 ]

[ "The", "improvement", "of", "our", "method", "is", "more", "prominent", "in", "the", "single-sentence", "classification", "task." ]

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

[ "M-IDPG-PHM", "is", "better", "than", "Compacter", "on", "four", "tasks", "and", "has", "the", "same", "performance", "on", "three", "tasks." ]

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

[ "Note", "that", "IDPG", "uses", "15K", "fewer", "parameters", "than", "the", "Compacter." ]

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

[ "However,", "the", "gap", "is", "mostly", "from", "RTE", "and", "QQP." ]

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

[ "(ii)", "Compared", "with", "other", "efficient", "transfer", "learning", "methods,", "IDPG", "performs", "slightly", "worse", "than", "the", "Compacter", "(Mahabadi", "et", "al.,", "2021)", "and", "Adapter", "(Houlsby", "et", "al.,", "2019),", "across", "the", "ten", "tasks." ]

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

[ "We", "observe", "that:", "(i)", "Our", "proposed", "method", "M-IDPG-PHM", "consistently", "outperforms", "the", "prompt", "tuning", "method", "and", "Ptuning", "v2", "by", "average", "3.1pt", "and", "1.6pt,", "respectively", "(except", "on", "the", "RTE", "dataset)." ]

[ 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0, 9, 0, 9, 0, 0, 0, 0, 13, 0 ]

[ "Table", "1", "shows", "the", "results", "of", "all", "the", "methods", "on", "full", "datasets", "across", "10", "NLU", "tasks." ]

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

[ "4.2", "Performance", "in", "high-resource", "scenario" ]

[ 0, 0, 0, 0, 0 ]

[ "Notably,", "Prompt-tuning134", "uses", "134", "prompt", "lengths", "in", "Table", "1,", "and", "it", "is", "set", "so", "to", "match", "the", "training", "parameters", "of", "the", "proposed", "method,", "M-IDPG-PHM." ]

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

[ "Additional", "training", "details", "can", "be", "found", "in", "Appendix", "A.1." ]

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

[ "For", "a", "fair", "comparison,", "all", "the", "pre-trained", "LMs", "are", "24-layer", "16-head", "RoBERTa-Large", "models", "(Liu", "et", "al.,", "2019)." ]

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

[ "Again,", "the", "difference", "between", "the", "first", "two", "is", "in", "the", "prompt", "generator,", "while", "M-IDPG-PHMGloVe", "uses", "GloVe", "to", "encode", "input", "sequences." ]

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

[ "We", "also", "explore", "three", "versions", "of", "multilayer", "instance-dependent", "generation", "methods:", "M-IDPG-DNN,", "M-IDPG-PHM,", "M-IDPG-PHM" ]

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

[ "The", "second", "one", "uses", "the", "PHM", "layer", "and", "only", "contains", "105K", "parameters." ]

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

[ "The", "first", "version", "is", "based", "on", "a", "2-layer", "perceptron", "generator,", "which", "contains", "1.5M", "parameters." ]

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

[ "We", "compare", "these", "against", "two", "versions", "of", "singlelayer", "instance-dependent", "generation", "methods:", "SIDPG-DNN", "and", "S-IDPG-PHM." ]

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

[ "Adapter-based", "fine-tuning:", "This", "efficient", "transfer", "learning", "method", "inserts", "an", "adaptation", "module", "inside", "each", "transformer", "layer", "including", "Compactor", "(Mahabadi", "et", "al.,", "2021)", "and", "Adapter", "(Houlsby", "et", "al.,", "2019)." ]

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

[ "Prompt", "tuning:", "We", "implemented", "two", "versions", "–", "standard", "prompt", "tuning", "(Lester", "et", "al.,", "2021)", "and", "multi-layer", "prompt", "tuning", "(Li", "and", "Liang,", "2021;", "Liu", "et", "al.,", "2021a)." ]

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

[ "Transformer", "fine-tuning:", "We", "instantiated", "two", "versions", "–", "a", "vanilla", "transformer", "fine-tuning", "(Liu", "et", "al.,", "2019)", "and", "the", "entailment-based", "finetuning", "(Wang", "et", "al.,", "2021)." ]

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

[ "We", "compare", "our", "proposed", "method", "with", "a", "wide", "range", "of", "methods,", "as", "follows:" ]

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

[ "SST-2,", "QNLI,", "RTE,", "MRPC,", "STS-B", "(Cer", "et", "al.,", "2017)", "and", "QQP." ]

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

[ "We", "evaluate", "on", "ten", "standard", "natural", "language", "understanding", "(NLU)", "datasets", "–", "MPQA", "(Wiebe", "et", "al.,", "2005),", "Subj", "(Pang", "and", "Lee,", "2004),", "CR", "(Hu", "and", "Liu,", "2004),", "MR", "(Pang", "and", "Lee,", "2005),", "and", "six", "tasks", "from", "GLUE", "(Wang", "et", "al.,", "2019),", "viz." ]

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

[ "4.1", "Experimental", "Setup" ]

[ 0, 0, 0 ]

[ "4", "Experiment", "Results" ]

[ 0, 0, 0 ]

[ "Method" ]

[ 0 ]

[ "For", "all", "the", "other", "tasks,", "we", "report", "accuracy." ]

[ 0, 0, 0, 0, 0, 0, 0, 7 ]

[ "We", "report", "the", "average", "of", "accuracy", "and", "F1", "for", "both", "MRPC", "and", "QQP,", "and", "average", "of", "Pearson", "and", "Spearman", "correlation", "coefficients", "for", "STS-B." ]

[ 0, 0, 0, 0, 0, 7, 0, 7, 0, 0, 13, 0, 13, 0, 0, 0, 7, 8, 8, 8, 8, 0, 13 ]

[ "Underline", "marks", "the", "best", "result", "among", "all", "prompt", "tuning", "methods." ]

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

[ "Bold", "marks", "the", "best", "result", "among", "all", "competing", "methods." ]

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

[ "We", "report", "average", "results", "across", "5", "runs", "with", "different", "initialization." ]

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

[ "Each", "methods", "are", "evaluated", "on", "full", "test", "sets", "(dev", "sets", "for", "GLUE", "tasks)." ]

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

[ "In", "short,", "assuming", "each", "layer", "generator", "Gi", "has", "form", "y", "=", "Wx", "+", "bi,", "we", "share", "the", "weight", "matrix", "Rm", "W", "across", "generators", "and", "set", "the", "bias", "term", "bi", "∈", "to", "be", "layer-specific,", "where", "i", "=", "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 ]

[ "However,", "simply", "generalizing", "our", "model", "(IDPG)", "to", "a", "multi-layer", "version", "(M-IDPG),", "will", "significantly", "increase", "the", "number", "of", "training", "parameters,", "since", "each", "layer", "requires", "an", "independent", "generator", "G.", "Instead,", "we", "explore", "different", "architectures", "in", "Section", "4.5.3", "to", "balance", "the", "number", "of", "tuned", "parameters", "against", "model", "performance." ]

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

[ "Following", "Prefix", "tuning", "(Li", "and", "Liang,", "2021)", "and", "P-tuning", "v2", "(Liu", "et", "al.,", "2021a),", "we", "prepend", "our", "generated", "prompts", "at", "each", "transformer", "layer", "to", "address", "the", "above", "issues." ]

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

[ "(ii)", "Generalizing", "to", "long", "sequence", "tasks:", "it", "is", "unclear", "that", "prompt", "tuning", "can", "perform", "well", "in", "tasks", "with", "long", "input", "when", "only", "a", "limited", "number", "of", "parameters", "can", "be", "inserted", "in", "single", "layer." ]

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

[ "While", "proven", "efficient", "in", "some", "specific", "settings,", "single", "layer", "prompt", "tuning", "has", "two", "main", "limitations:", "(i)", "Capturing", "deep", "contextual", "information:", "the", "impact", "of", "the", "first-layer", "prompts", "on", "final", "prediction", "is", "low", "when", "transformer", "goes", "deeper." ]

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