Datasets:

havens2

/

naacl2022

like 0

[ "For", "sentence-pair", "tasks,", "we", "hypothesize", "that", "inserting", "prompt", "into", "position", "1", "can", "better", "align", "the", "two", "input", "sentences." ]

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

[ "This", "result", "is", "intuitive", "for", "single-sentence", "tasks", "since", "prompt", "in", "position", "0", "can", "be", "regarded", "as", "the", "premise", "and", "original", "input", "sentence", "as", "the", "hypothesis." ]

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

[ "Looking", "at", "the", "prompt", "position", "in", "downstream", "tasks", "first,", "Figure", "7(a)", "shows", "that", "for", "both", "standard", "prompt", "tuning", "and", "our", "proposed", "method,", "the", "best", "position", "is", "0", "for", "single-sentence", "tasks", "and", "1", "for", "sentence-pair", "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, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "We", "conduct", "a", "comprehensive", "study", "of", "the", "prompt", "position", "for", "our", "proposed", "method", "in", "both", "supplementary", "training", "and", "downstream", "fine-tuning", "phases." ]

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

[ "A.5.2", "Prompt", "Position", "As", "we", "discussed", "in", "Section", "A.4,", "the", "prompt", "position", "has", "a", "direct", "impact", "on", "the", "prediction", "results." ]

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

[ "One", "surprising", "result", "is", "that", "the", "mixed", "model", "of", "Residual", "and", "LayerNorm", "has", "significantly", "poorer", "performance." ]

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

[ "We", "observe", "that", "adding", "LayerNorm", "slightly", "improves", "the", "voting", "results,", "while", "residual", "performs", "slightly", "worse." ]

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

[ "Note", "that,", "to", "balance", "the", "token", "embedding", "and", "sentence", "embedding,", "we", "apply", "LayerNorm", "to", "each", "embedding", "first,", "then", "after", "the", "add-up,", "use", "LayerNorm", "again", "to", "control", "the", "generated", "tokens." ]

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

[ "et", "al.,", "2016)", "is", "also", "added", "to", "normalize", "the", "generated", "token", "embedding;", "(iii)", "residual", "+", "layerNorm:", "a", "mixed", "model", "that", "uses", "both", "the", "residual", "component", "and", "LayerNorm." ]

[ 0, 0, 0, 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", "explore", "three", "different", "architectures", "for", "the", "proposed", "PHM-based", "generator:", "(i)", "Residual:", "a", "residual", "structure", "(He", "et", "al.,", "2016)", "is", "applied", "to", "add", "the", "sentence", "representation", "to", "each", "generated", "tokens;", "(ii)", "LayerNorm:", "layer", "normalization", "(Ba" ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 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.5.1", "Generator", "Architecture", "Exploration" ]

[ 0, 0, 0, 0 ]

[ "A.5", "Ablation", "study", "for", "single-layer", "IDPG" ]

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

[ "We", "report", "average", "results", "and", "voting", "results", "across", "5", "runs." ]

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

[ "We", "double", "the", "MNLI", "dataset", "by", "reordering", "the", "two", "sentences", "on", "one", "shard,", "and", "use", "the", "doubled", "dataset", "during", "intermediate", "training." ]

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

[ "We", "further", "reduce", "the", "distribution", "difference", "by", "reconstructing", "the", "supplementary", "training", "data." ]

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

[ "As", "shown", "in", "Figure", "6,", "we", "test", "five", "different", "insertion", "positions", "(Pos", "0–4)", "for", "sentence", "pair", "tasks", "and", "three", "different", "positions", "(Pos", "0,", "1,", "4)", "for", "single", "sentence", "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, 0, 0 ]

[ "In", "this", "way,", "the", "length", "difference", "in", "MNLI", "and", "MRPC", "becomes", "more", "balanced:", "4.8", "vs.", "0.6", "+", "5", "=", "5.6." ]

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

[ "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 ]

[ "Then,", "when", "fine-tuning", "downstream", "tasks" ]

[ 0, 0, 0, 0, 0 ]

[ "For", "example,", "assuming", "that", "we", "are", "adding", "a", "prompt", "with", "a", "length", "t", "=", "5", "after", "the", "second", "sentence", "in", "the", "supplementary", "training", "stage", "on", "MNLI." ]

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

[ "One", "natural", "solution", "to", "smooth", "the", "length", "distribution", "difference", "between", "tasks", "is", "to", "insert", "prompt", "in", "both", "supplementary", "training", "and", "downstream", "finetuning", "stage." ]

[ 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,", "the", "length", "of", "the", "first", "sentence", "in", "MNLI", "is", "9.8", "longer", "than", "the", "second", "sentence", "on", "average,", "while", "this", "length", "difference", "in", "MRPC", "is", "only", "0.6." ]

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

[ "Figure", "5", "provides", "a", "comprehensive", "statistic", "among", "all", "sentence", "pair", "tasks", "in", "GLUE", "benchmark." ]

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

[ "However,", "a", "drawback", "of", "supplementary", "training", "is", "that", "if", "the", "data", "distribution", "of", "the", "downstream", "tasks", "is", "quite", "different", "from", "the", "supplementary", "training", "task,", "i.e.,", "MRPC", "vs.", "MNLI", "(Wang", "et", "al.,", "2019),", "it", "may", "harm", "the", "downstream", "performance." ]

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

[ "Following", "this", "idea,", "we", "also", "conduct", "intermediate", "training", "for", "single-layer", "IDPG." ]

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

[ "According", "to", "previous", "works", "(Phang", "et", "al.,", "2018;", "Wang", "et", "al.,", "2021),", "supplementing", "pre-trained", "LMs", "with", "rich", "data", "helps", "tasks", "with", "limited", "labels", "and", "stabilizes", "downstream", "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, 0, 0, 0 ]

[ "A.4", "Supplementary", "Training", "for", "Single-layer" ]

[ 0, 0, 0, 0, 0 ]

[ "While", "in", "Table", "1,", "we", "tune", "the", "learning", "4", "4,", "1e−", "3,", "5e−", "to", "make", "the", "rate", "from", "}", "fair", "comparison", "with", "other", "models." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 5, 6, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "M-IDPG-PHM),", "this", "is", "because", "we", "tune", "the", "learning", "rate", "more", "carefully", "in", "Table", "6", "(lr", "2,", "7e−", "4,", "3,", "1e−", "1e−", "{", "4", "1e−", ")", "to", "seek", "the", "best", "performance", "each", "model", "can", "reach." ]

[ 1, 0, 0, 0, 0, 0, 0, 3, 4, 0, 0, 0, 0, 0, 3, 5, 5, 6, 5, 5, 5, 0, 5, 5, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "Note", "that", "the", "M", "version", "model", "with", "m", "=", "16", "and", "previous", "layer", "as", "input", "one", "is", "slightly", "higher", "than", "the", "results", "shown", "in", "Table", "1(Cf." ]

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

[ "We", "provide", "a", "detailed", "result", "table", "for", "all", "compared", "methods", "in", "Section", "4.5.3." ]

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

[ "A.3", "Detailed", "results", "for", "Multi-layer", "Architecture", "Exploration" ]

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

[ "We", "provide", "a", "detailed", "information", "in", "Table", "5", "for", "10", "NLU", "datasets", "we", "used." ]

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

[ "A.2", "Datasets" ]

[ 0, 0 ]

[ "Total", "parameters:", "1024", "×", "24", "=", "216K.", "1024" ]

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

[ "M-IDPG-DNN:", "hidden", "size", "d", "=", "1024,", "generator", "hidden", "size", "m", "=", "16,", "prompt", "length", "t", "=", "5." ]

[ 1, 3, 4, 3, 0, 5, 3, 4, 4, 3, 5, 6, 3, 4, 3, 5, 6 ]

[ "First", "PHM", "layer", "16", "+", "16", "=", "1K", "paramW1", "takes", "1024/16", "eters,", "second", "PHM", "layer", "W2", "takes", "16/16", "1024/16", "1024", "the", "shared", "matrix", "Ai", "takes", "16316", "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, 0, 0 ]

[ "=", "5,", "user", "defined", "n", "=", "16", "(Cf.", "Equation", "4)." ]

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

[ "M-IDPG-PHM:", "hidden", "size", "d", "=", "1024,", "generator", "hidden", "size", "m", "=", "16,", "prompt", "length", "t" ]

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

[ "First", "PHM", "4", "+", "16", "=", "1216", "palayer", "W1", "takes", "300/4", "×", "rameters,", "second", "PHM", "layer", "W2", "takes", "16/4", "1024/4", "1024", "the", "shared", "matrix", "Ai", "takes", "43", "parameters:", "141K." ]

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

[ "×", "M-IDPG-PHM-GloVe:", "input", "vector", "size", "300,", "generator", "hidden", "size", "m", "=", "16,", "prompt", "length", "t", "=", "5,", "user", "defined", "n", "=", "4", "(Cf.", "Equation", "4)." ]

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

[ "S-IDPG-DNN:", "hidden", "size", "d", "=", "1024,", "generator", "hidden", "size", "m", "=", "256,", "prompt", "length", "t", "=", "5." ]

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

[ "First", "PHM", "layer", "W1", "takes", "1024/16", "16", "+", "256", "=", "16.25K", "256/16", "parameters,", "second", "PHM", "layer", "W2", "takes", "256/16", "×", "1024", "=", "85K", "parameters,", "5", "×", "the", "shared", "matrix", "Ai", "takes", "163", "=", "4K", "(Note", "we", "use", "one", "shared", "matrix", "in", "single", "version", "IDPG)." ]

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

[ "P-tuning", "v2:", "prompt", "length", "t", "=", "5,", "inserted", "lay1024", "=", "120K.", "S-IDPG-PHM:", "hidden", "size", "d", "=", "1024,", "generator", "hidden", "size", "m", "=", "256,", "prompt", "length", "t", "=", "5,", "user", "defined", "n", "=", "16", "(Cf.", "Equation", "4)." ]

[ 1, 2, 3, 4, 3, 0, 5, 0, 0, 0, 0, 1, 3, 4, 3, 0, 5, 3, 4, 4, 3, 0, 5, 3, 4, 3, 0, 5, 0, 0, 3, 0, 5, 0, 0, 0 ]

[ "Prompt-tuning-134:", "prompt", "length", "t", "=", "134." ]

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

[ "Prompt-tuning:", "prompt", "length", "t", "=", "5." ]

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

[ "Total", "parameters:", "(1024", "2", "=", "1.55M.", "16" ]

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

[ "size", "m", "=", "16." ]

[ 0, 3, 0, 5 ]

[ "Adapter:", "hidden", "size", "d", "=", "1024,", "adapter", "hidden", "16", "+", "16", "+" ]

[ 1, 3, 4, 3, 0, 5, 1, 2, 5, 0, 0, 0 ]

[ "24", "4" ]

[ 0, 0 ]

[ "Down-project", "2", "=", "48K,", "downsi", "matrix", "takes", "1024/4", "4", "×", "project", "ti", "matrix", "takes", "16/4", "2", "=", "0.75K,", "24", "×", "×", "2", "=", "0.75K,", "up-project", "hidden", "bias", "takes", "16", "24", "si", "and", "ti", "matrix", "takes", "the", "same", "number", "of", "parameters", "as", "down-projector,", "the", "output", "bias", "takes", "2", "=", "48K,", "the", "shared", "matrix", "Ai", "takes", "1024", "×", "43", "2", "=", "3K.", "Total", "parameters:", "48", "+", "0.75", "+", "0.75", "+", "48", "+", "0.75", "+", "48", "+", "3", "=", "149.25K." ]

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

[ "Compacter:", "hidden", "size", "d", "=", "1024,", "adapter", "hidden", "size", "m", "=", "16,", "user", "defined", "n", "=", "4,", "each", "transformer", "layer", "inserts", "2", "compacters." ]

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

[ "We", "report", "the", "detailed", "model", "hyperparameters", "for", "each", "method", "in", "Table", "1", "and", "illustrate", "how", "numbers", "in", "Table", "2", "are", "computed." ]

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

[ "A.1.2", "Model", "hyperparameters" ]

[ 0, 0, 0 ]

[ "For", "all", "adapter-based", "and", "prompt-based", "methods,", "we", "train", "them", "more", "sufficiently", "(with", "fifty", "epochs)", "on", "small", "datasets", "(i.e.,", "MPQA,", "Subj,", "CR,", "MR,", "RTE,", "MRPC,", "STS-B)." ]

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

[ "Note", "that", "for", "both", "transformer", "fine-tuning", "methods,", "including", "RoBERTa", "(Liu", "et", "al.,", "2019)", "and", "EFL", "(Wang", "et", "al.,", "2021),", "we", "follow", "their", "official", "training", "instructions,", "i.e.,", "using", "a", "polynomial", "learning", "rate", "scheduler", "with", "6%", "steps", "to", "warm", "up", "and", "tuning", "for", "ten", "epochs." ]

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

[ "The", "detailed", "model", "hyperparameters", "are", "listed", "in", "Table", "4." ]

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

[ "We", "use", "RoBERTa-Large", "(Liu", "et", "al.,", "2019)", "model", "implemented", "by", "Fairseq", "(Ott", "et", "al.,", "2019)", "as", "our", "basic", "model." ]

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

[ "A", "Appendix" ]

[ 0, 0 ]

[ "References" ]

[ 0 ]

[ "Acknowledgment" ]

[ 0 ]

[ "This", "work", "provided", "a", "new", "research", "angle", "for", "prompt-tuning", "of", "a", "pre-trained", "language", "model." ]

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

[ "It", "is", "also", "on", "par", "with", "the", "lightweight", "adapter", "tuning", "methods", "such", "as", "Compacter", "while", "using", "a", "similar", "amount", "of", "trainable", "parameters." ]

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

[ "Despite", "adding", "fewer", "parameters", "than", "prompt", "tuning,", "IDPG", "shows", "consistent", "improvement." ]

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

[ "Parameterized", "Hypercomplex", "Multiplication", "(PHM)", "is", "applied", "to", "shrink", "the", "training", "parameters", "in", "our", "prompt", "generator,", "which", "helps", "us", "build", "an", "extreme", "lightweight", "generation", "model." ]

[ 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", "method", "first", "factors", "in", "an", "instance-dependent", "prompt,", "which", "is", "robust", "to", "data", "variance." ]

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

[ "We", "have", "introduced", "IDPG,", "an", "instance-dependent", "prompt", "generation", "model", "that", "generalizes", "better", "than", "the", "existing", "prompt", "tuning", "methods." ]

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

[ "6", "Conclusion", "and", "Discussion" ]

[ 0, 0, 0, 0 ]

[ "Our", "proposed", "method,", "IDPG,", "is", "the", "first", "prompt", "generator", "that", "is", "not", "only", "taskspecific", "but", "also", "instance-specific." ]

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

[ "All", "these", "methods", "focus", "on", "task-specific", "prompt", "optimization." ]

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

[ "Two", "contemporaneous", "works", "–", "prompt", "tuning", "(Lester", "et", "al.,", "2021)", "and", "P-tuning", "(Liu", "et", "al.,", "2021b),", "interleave", "the", "training", "parameters", "in", "the", "input", "embedding", "layer", "instead", "of", "each", "transformer", "layer." ]

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

[ "Prefix", "tuning", "(Li", "and", "Liang,", "2021)", "and", "P-tuningv2", "(Liu", "et", "al.,", "2021a)", "prepend", "a", "sequence", "of", "trainable", "embeddings", "at", "each", "transformer", "layer", "and", "optimizes", "them." ]

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

[ "Recently,", "several", "researchers", "have", "proposed", "continuous", "prompts", "training", "to", "overcome", "the", "challenges", "in", "discrete", "prompt", "searching." ]

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

[ "Existing", "works", "including", "LMBFF", "(Gao", "et", "al.,", "2021a;", "Wang", "et", "al.,", "2021)", "explored", "the", "prompt", "searching", "in", "a", "few-shot", "setting." ]

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

[ "Prompting:", "Hand-crafted", "prompts", "were", "shown", "to", "be", "helpful", "to", "adapt", "generation", "in", "GPT-3", "(Brown", "et", "al.,", "2020)." ]

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

[ "et", "al.,", "2021)", "substitutes", "the", "down-projector", "and", "upprojector", "matrices", "by", "a", "sum", "of", "Kronecker", "products,", "reducing", "the", "parameters", "by", "a", "large", "margin", "while", "maintaining", "the", "overall", "performance." ]

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

[ "On", "the", "GLUE", "benchmark,", "adapters", "attain", "within", "0.4%", "of", "the", "performance", "of", "full", "fine-tuning", "by", "only", "training", "3.6%", "parameters", "per", "task." ]

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

[ "Pfeiffer", "et", "al.,", "2020),", "in", "which", "adapter", "layers", "–", "small", "bottleneck", "layers", "–", "are", "inserted", "and", "trained", "between", "frozen", "pre-trained", "transformer", "layers." ]

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

[ "Adapter", "tuning", "has", "emerged", "as", "a", "novel", "parameter-efficient", "transfer", "learning", "paradigm", "(Houlsby", "et", "al.,", "2019;" ]

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

[ "Adapter", "Tuning:" ]

[ 0, 0 ]

[ "More", "recently,", "EFL", "(Wang", "et", "al.,", "2021)", "proposed", "a", "task", "transformation", "paradigm,", "improving", "single", "sentence", "tasks", "with", "less", "labels", "using", "rich", "sentence-pair", "datasets." ]

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

[ "All", "of", "them", "applied", "pre-fine", "tuning", "on", "NLI", "datasets." ]

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

[ "A", "series", "of", "work", "(SentenceBERT", "(Reimers", "and", "Gurevych,", "2019),", "BERT-flow", "(Li", "et", "al.,", "2020),", "SimCSE", "(Gao", "et", "al.,", "2021b))", "explored", "intermediate", "training", "to", "improve", "STS", "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, 13, 0 ]

[ "Existing", "works", "(Phang", "et", "al.,", "2018;", "Liu", "et", "al.,", "2019)", "have", "observed", "that", "starting", "from", "the", "fine-tuned", "MNLI", "model", "results", "in", "a", "better", "performance", "than", "directly", "from", "the", "vanilla", "pre-trained", "models", "for", "RTE,", "STS,", "and", "MRPC", "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, 0, 0, 0, 0, 0, 13, 13, 0, 13, 0 ]

[ "Supplementary", "Training:" ]

[ 0, 0 ]

[ "5", "Related", "Work" ]

[ 0, 0, 0 ]

[ "In", "general,", "the", "performance", "of", "IDPG", "in", "downstream", "tasks", "improves", "gradually", "when", "using", "a", "larger", "prompt", "length", "(Cf.", "Appendix", "A.7)." ]

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

[ "4.5.6", "IDPG", "Scalability", "We", "study", "our", "proposed", "model’s", "scalability", "in", "this", "section." ]

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

[ "As", "shown", "in", "Figure", "4,", "prompts", "bring", "the", "similar", "sentences", "closer", "while", "pushing", "the", "dissimilar", "ones", "apart." ]

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

[ "Both", "models", "are", "finetuned", "on", "STS-B", "training", "set." ]

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

[ "We", "compare", "a", "vanilla", "model", "without", "any", "prompts", "with", "M-IDPG-PHM." ]

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

[ "More", "results", "can", "be", "found", "in", "Appendix", "A.6." ]

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

[ "We", "sort", "all", "sentence", "pairs", "in", "STS-B", "dev", "set", "in", "descending", "order", "by", "the", "cosine" ]

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

[ "Given", "two", "sentences,", "we", "encode", "each", "of", "them", "by", "one", "of", "the", "comparison", "models", "and", "compute", "the", "cosine", "similarity." ]

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

[ "4.5.5", "How", "Prompts", "Help?" ]

[ 0, 0, 0, 0 ]

[ "An", "interesting", "finding", "is", "that", "the", "impact", "of", "supplementary", "training", "on", "S-IDPG", "is", "high", "while", "it", "is", "limited", "for", "M-IDPG." ]

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