Datasets:

havens2

/

naacl2022

like 0

[ "For", "CLEVR,", "the", "image", "representation", "is", "extracted", "from", "a", "pretrained", "ResNet50", "and", "projected", "into", "a", "tensor", "of", "size", "(32,7,7)", "before", "being", "flattened." ]

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

[ "the", "answer", "embedding", "of", "dimension", "32", "(resp.", "128),", "and", "the", "image", "representation." ]

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

[ "At", "every", "time", "step,", "the", "LSTM", "input", "is", "then", "the", "concatenation", "of", "the", "word", "embedding", "of", "dimension", "32", "(resp." ]

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

[ "256)", "units", "for", "the", "policy", "network." ]

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

[ "VQAv2),", "we", "used", "a", "single-layer", "LSTM", "with", "64", "(resp." ]

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

[ "For", "CLEVR", "(resp." ]

[ 0, 13, 0 ]

[ "A.4", "Language", "Agent", "Networks", "and", "Training" ]

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

[ "10)", "words." ]

[ 5, 3 ]

[ "Finally,", "questions", "are", "limited", "to", "20", "(resp." ]

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

[ "Besides,", "we", "uniformly", "sample", "the", "answer", "in", "the", "set", "of", "reference", "answers", "for", "each", "image", "to", "reduce", "the", "bias", "in", "the", "distribution", "of", "answers." ]

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

[ "20k)", "images", "of", "the", "validation", "set." ]

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

[ "all", "the", "images)", "of", "the", "training", "dataset,", "and", "is", "then", "evaluated", "on", "the", "first", "5k", "(resp." ]

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

[ "100k)", "episodes", "over", "the", "first", "20k", "images", "(resp." ]

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

[ "VQAv2),", "the", "RL", "language", "agent", "is", "trained", "for", "50k", "(resp." ]

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

[ "For", "CLEVR", "(resp." ]

[ 0, 13, 0 ]

[ "A.3", "Dataset", "split" ]

[ 0, 0, 0 ]

[ "A.2", "Answer", "filtering", "For", "each", "dataset,", "we", "remove", "yes", "and", "no", "question-answer", "pairs", "which", "frequency", "largely", "exceeds", "other", "answers,", "to", "avoid", "any", "bias", "in", "the", "question", "generation", "process,", "as", "usually", "done", "in", "the", "VQG", "litterature", "(Mostafazadeh", "et", "al.,", "2016)." ]

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

[ "For", "the", "CIDEr", "score,", "we", "used", "the", "nlg-eval", "implementation6." ]

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

[ "A.1", "Evaluation", "Metrics", "For", "the", "BLEU", "and", "METEOR", "scores,", "we", "used", "the", "NLTK5", "implementations", "with", "the", "smoothing", "function", "number", "2", "for", "the", "BLEU", "score." ]

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

[ "A", "Dataset", "and", "training", "details" ]

[ 0, 0, 0, 0, 0 ]

[ "References" ]

[ 0 ]

[ "Acknowledgements" ]

[ 0 ]

[ "Although", "it", "comes", "with", "its", "limitations,", "the", "truncated", "RL", "algorithm", "provided", "by", "TrufLL", "gets", "free", "from", "labelled", "data", "in", "task-oriented", "language", "models,", "presents", "interesting", "language", "generation", "properties,", "and", "provides", "a", "generic", "and", "transferable", "method", "to", "learn", "any", "NLG", "problem." ]

[ 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, 11, 0 ]

[ "To", "our", "knowledge,", "this", "is", "the", "first", "RL-based", "algorithm", "dedicated", "to", "learning", "a", "word-based", "text-generation", "task,", "which", "does", "not", "rely", "on", "a", "pre-training", "phase", "while", "scaling", "to", "large", "vocabularies." ]

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

[ "We", "proposed", "TrufLL,", "an", "original", "approach", "to", "learn", "a", "natural", "language", "generation", "(NLG)", "task", "using", "RL,", "without", "the", "usual", "pre-training", "phase", "requiring", "supervised", "datasets." ]

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

[ "Yet,", "they", "still", "face", "optimization", "and", "computational", "challenges", "(Parisotto", "et", "al.,", "2020)." ]

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

[ "Garg", "et", "al.,", "2021)", "finetune", "pretrained", "GPT-2", "models", "with", "RL", "for", "language", "generation", "tasks", "without", "task-related", "data,", "only", "using", "reward", "signals." ]

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

[ "Learning", "Language", "Models", "from", "scratch." ]

[ 0, 0, 0, 0, 0 ]

[ "Besides,", "CALM", "is", "only", "evaluated", "on", "a", "vocabulary", "of", "697", "tokens,", "and", "on", "4-words", "action", "sequences." ]

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

[ "Yet,", "its", "truncation", "language", "model", "remains", "fine-tuned", "on", "the", "RL", "dataset." ]

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

[ "Similarly", "to", "TrufLL,", "CALM", "(Yao", "et", "al.,", "2020)", "combines", "a", "pretrained", "language", "model", "to", "prune", "the", "action", "space", "with", "a", "DeepQ", "network,", "aka", "DRNN", "(He", "et", "al.,", "2016)." ]

[ 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 ]

[ "Closer", "to", "our", "work,", "a", "few", "algorithms", "(Ammanabrolu", "and", "Riedl,", "2018)", "use", "the", "structure", "of", "language", "to", "prune", "the", "action", "space", "of", "text-based", "games,", "but", "within", "value-based", "algorithms,", "which", "are", "less", "scalable", "to", "large", "vocabularies." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 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)", "proposes", "Q-learning", "algorithms", "with", "an", "elimination", "signal", "to", "eliminate", "forbidden", "actions." ]

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

[ "RL", "methods", "for", "Language", "Action", "Spaces." ]

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

[ "Several", "RL", "algorithms", "have", "been", "developed", "to", "tackle", "large", "discrete", "action", "spaces." ]

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

[ "Hence,", "Dulac-Arnold", "et", "al.", "(2015);", "Tennenholtz", "and", "Mannor", "(2019);", "Chandak", "et", "al.", "(2019)", "embed", "the", "actions", "into", "a", "continuous", "action", "space,", "and", "then", "use", "classic", "RL", "algorithms", "to", "learn", "a", "policy", "over", "this", "continuous", "space." ]

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

[ "Bahdanau", "et", "al.", "(2016);", "Rennie", "et", "al.", "(2017)", "use", "RL", "to", "train", "language", "models", "as", "an", "alternative", "to", "SL", "to", "prevent", "typical", "text", "degeneration", "issues,", "but", "within", "training", "algorithms", "relying", "on", "ground-truth", "examples", "from", "labelled", "datasets." ]

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

[ "Yet,", "the", "former", "uses", "slot", "filling", "with", "template", "questions,", "while", "the", "later", "computes", "a", "mixed", "objective", "with", "a", "MLE", "loss", "using", "ground-truth", "sentences." ]

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

[ "fine-tuning." ]

[ 0 ]

[ "Following", "(Singh", "et", "al.,", "2002;", "Lemon", "and", "Pietquin,", "2007),", "recent", "RL-based", "taskoriented", "dialogues", "(De", "Vries", "et", "al.,", "2017;", "Das", "et", "al.,", "2017;", "Lewis", "et", "al.,", "2017;", "Narasimhan", "et", "al.,", "2015)", "have", "been", "developed,", "where", "the", "policy", "language", "model", "is", "generally", "pretrained", "with", "SL", "followed", "RL" ]

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

[ "RL", "and", "NLP", "Tasks." ]

[ 0, 0, 0, 0 ]

[ "6", "Related", "work" ]

[ 0, 0, 0 ]

[ "The", "context", "can", "be", "any", "kind", "of", "data", "structure", "if", "it", "is", "a", "(natural", "language,", "database,", "video,", "etc):", "linguistic", "input,", "TrufLL", "can", "be", "applied", "for", "instance", "to", "text", "summarization,", "paraphrase", "generation", "(with", "reward", "functions", "based", "on", "similarity", "scores", "between", "the", "context", "and", "the", "generated", "language)", "or", "text-based", "games", "(Ammanabrolu", "and", "Riedl,", "2018)." ]

[ 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, 0, 0 ]

[ "Reward", "functions", "for", "such", "tasks", "can", "be", "based", "on", "similarity", "scores", "between", "the", "generated", "language", "and", "the", "associated", "image", "or", "image", "region,", "which", "can", "be", "computed", "using", "pretrained", "language", "representations", "such", "as", "BERT", "(Devlin", "et", "al.,", "2019)", "or", "multi-modal", "pretrained", "systems", "such", "as", "ViLBERT", "(Lu", "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, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "Other", "interesting", "tasks", "for", "TrufLL", "include", "the", "ones", "typically", "found", "in", "Vision", "and", "Language", "Representation", "Learning", "(Lu", "et", "al.,", "2020a),", "such", "as", "Image", "Captioning,", "Grounding", "Referring", "Expressions", "(generation", "of", "a", "referring", "expression", "over", "a", "specific", "bounding", "box", "of", "an", "image),", "Captionbased", "Image", "Retrieval", "(generation", "of", "a", "caption", "that", "discriminates", "an", "image", "between", "a", "set", "of", "images)." ]

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

[ "Additionally,", "the", "RL", "framework", "allows", "to", "optimize", "non-differentiable", "objectives,", "making", "TrufLL", "a", "natural", "choice", "to", "learn", "end-to-end", "task-oriented", "dialogs,", "such", "as", "(De", "Vries", "et", "al.,", "2017;", "Das", "et", "al.,", "2017)." ]

[ 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 ]

[ "Hence,", "the", "method", "is", "transferable", "to", "a", "wide", "variety", "of", "NLG", "tasks,", "without", "requiring", "upfront", "large-scale", "labelled", "datasets." ]

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

[ "TrufLL", "learns", "conditional", "language", "models", "able", "to", "solve", "specific", "Natural", "Language", "Generation", "tasks", "given", "a", "context", "c.", "For", "solving", "such", "tasks,", "it", "only", "requires", "the", "context,", "a", "reward", "function", "that", "scores", "the", "language", "generated", "by", "the", "RL", "agent", "with", "respect", "to", "the", "task,", "and", "eventually", "a", "few", "natural", "language", "demonstrations", "fed", "as", "input", "prompt", "to", "the", "generic", "language", "model", "used", "in", "the", "truncation", "algorithm." ]

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

[ "Generalization", "of", "the", "approach." ]

[ 0, 0, 0, 0 ]

[ "Finally,", "we", "explore", "TrufLL", "with", "a", "pre-training", "phase", "in", "Table", "10." ]

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

[ "In", "VQA,", "we", "apply", "temperature", "scheduling", "on", "the", "LM", "to", "perform", "fine-grained", "truncations", "in", "Table", "9", "of", "B.2." ]

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

[ "In", "Table", "8,", "we", "observe", "that", "rewarding", "an", "agent", "with", "a", "BLEU", "score", "is", "sub-optimal", "in", "both", "language", "and", "task", "scores", "on", "CLEVR." ]

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

[ "We", "sweep", "over", "truncation", "hyper-parameters", "in", "Table", "6", "of", "Appendix", "B." ]

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

[ "Additional", "experiments." ]

[ 0, 0 ]

[ "We", "leave", "such", "possibilities", "to", "future", "works." ]

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

[ "It", "could", "be", "improved", "with", "regularisation", "techniques", "and", "the", "use", "of", "TruFLL", "within", "state-of-the-art", "off-policy", "RL", "algorithms." ]

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

[ "However,", "the", "approach", "does", "not", "manage", "to", "sufficiently", "scale", "to", "VQAv2." ]

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

[ "As", "its", "sumVA", "ratios", "are", "very", "close", "to", "1,", "the", "agent", "has", "learned", "to", "generalize", "over", "the", "full", "vocabulary." ]

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

[ "On", "CLEVR,", "the", "TrufLLoff", "has", "lower", "-", "yet", "close", "-", "performance", "on", "language", "and", "task", "scores", "than", "TrufLL." ]

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

[ "Method" ]

[ 0 ]

[ "Thus,", "we", "need", "to", "unbiased", "the", "PG", "by", "using", "an", "importance", "sampling", "term", "between", "the", "exploratory", "policy", "π−θ", "and", "the", "behavior", "policy", "πθ", "(Degris", "et", "al.," ]

[ 0, 0, 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", "such", "a", "setting,", "we", "adopt", "an", "off-policy", "training", "scheme,", "where", "the", "trajectories", "used", "to", "learn", "the", "behavior", "πθ", "at", "training", "time", "are", "sampled", "under", "a", "different", "policy,", "the", "truncated", "policy", "π−θ", "." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 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", "section,", "we", "investigate", "if", "we", "can", "directly", "learn", "a", "policy", "over", "the", "full", "vocabulary,", "and", "thus", "removing", "the", "truncation", "at", "test", "time." ]

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

[ "Hence,", "the", "algorithm", "requires", "the", "truncation,", "and", "a", "fortiori", "the", "language", "model,", "at", "test", "time." ]

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

[ "So", "far,", "TrufLL", "directly", "learns", "the", "truncated", "policy", "over", "the", "truncated", "vocabulary", "in", "an", "on-policy", "scheme." ]

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

[ "Removing", "the", "truncation", "at", "evaluation", "with", "offpolicy", "RL." ]

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

[ "5.3", "Discussion" ]

[ 0, 0 ]

[ "As", "PG", "solely", "optimizes", "the", "task", "success", "ratio,", "this", "may", "reduce", "overall", "language", "quality,", "the", "re-ranking", "thus", "retrieves", "the", "best", "syntactically", "sentences", "a", "posteriori." ]

[ 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", "greedy", "decoding", "produces", "the", "best", "outcome", "for", "pretrained", "models,", "lm-ranking", "provides", "an", "excellent", "trade-off", "between", "task", "performance", "and", "language", "quality", "with", "RL-based", "methods." ]

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

[ "In", "Table", "4,", "we", "evaluate", "the", "Decoding", "procedure:", "text", "sampling", "procedures", "described", "in", "Section", "4.5." ]

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

[ "It", "confirms", "that", "TrufLL", "(Ext-LM)", "could", "be", "an", "alternative", "approach", "as", "it", "has", "an", "excellent", "trade-off", "between", "language", "quality,", "diversity,", "and", "grounding." ]

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

[ "Although", "its", "questions", "are", "less", "grounded,", "they", "are", "diverse,", "which", "suggests", "that", "they", "follow", "a", "different", "distribution", "from", "the", "initial", "VQA", "dataset." ]

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

[ "Finally,", "unlike", "TrufLL", "(Task-LM)", "which", "suffers", "from", "syntactic", "errors,", "TrufLL", "(Ext-LM)", "produces", "language", "that", "qualitatively", "competes", "with", "pretrain", "models", "(53%),", "with", "a", "similar", "ratio", "of", "syntactic", "uncorrect", "samples." ]

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

[ "Yet,", "they", "exhibit", "significantly", "less", "diversity", "with", "the", "reference", "language;", "this", "suggests", "in", "particular", "that", "pretrain+RL", "fails", "to", "go", "beyond", "the", "initial", "task-data", "distribution." ]

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

[ "supervised", "baselines", "produce", "the", "best", "language,", "while", "being", "accurately", "grounded." ]

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

[ "Yet,", "it", "fails", "to", "generate", "correct", "and", "grounded", "language;", "it", "is", "thus", "not", "a", "viable", "approach", "despite", "its", "diverse", "In", "line", "with", "the", "automatic", "metrics,", "the", "output." ]

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

[ "Among", "the", "RL", "from", "scratch", "baselines,", "we", "selected", "scratch+KL-task", "as", "the", "only", "model", "producing", "sometimes", "meaningful", "questions." ]

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

[ "Human", "Evaluation:", "Figure", "3", "details", "the", "Human", "Evaluation", "results." ]

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

[ "questions", "which", "differs", "from", "the", "VQA", "distribution,", "yet", "consistent", "with", "the", "context." ]

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

[ "Pretrained", "baselines", "tend", "to", "output", "a", "question", "closer", "to", "the", "reference", "question", "whereas", "TrufLL", "outputs", "original" ]

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

[ "TrufLL", "and", "the", "pretrained", "baselines", "successfully", "generate", "a", "question", "giving", "the", "expected", "answer", "(\"Black\"),", "while", "the", "RL", "from", "scratch", "baselines", "fail,", "and", "even", "showcase", "degenerated", "language." ]

[ 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 ]

[ "the", "poor", "diversity", "behavior", "This", "suggests", "that", "observed", "on", "CLEVR", "is", "likely", "attributable", "to", "the", "small", "vocabulary", "and", "synthetic", "language", "distribution." ]

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

[ "Interestingly,", "TrufLL", "displays", "a", "self-BLEU", "score", "similar", "to", "the", "pretrained", "baselines." ]

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

[ "The", "similarity", "between", "TrufLL", "(Task-LM)", "and", "TrufLL", "(Ext-LM)", "results", "suggests", "that", "the", "truncation", "approach", "is", "viable", "when", "using", "a", "generic", "LM", "whose", "original", "vocabulary", "distribution", "differs", "from", "the", "task." ]

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

[ "Although", "TrufLL", "does", "not", "outperform", "the", "performance", "of", "the", "pretrained", "baselines", "anymore,", "it", "still", "leads", "to", "similar", "performances,", "and", "satisfactory", "language", "scores." ]

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

[ "First,", "the", "scratch", "baselines", "keep", "failing", "to", "learn", "a", "valuable", "policy,", "with", "performance", "scores", "and", "n-grams", "metrics", "close", "to", "zero." ]

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

[ "Quantitative", "performance:", "Table", "3", "reports", "the", "VQAv2", "results,", "for", "which", "TrufLL", "and", "the", "baselines", "present", "a", "similar", "trend", "than", "on", "CLEVR." ]

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

[ "We", "scale", "our", "approach", "to", "natural", "language", "with", "large", "vocabulary", "(15k", "tokens)", "through", "the", "VQAv2", "dataset." ]

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

[ "In", "CLEVR,", "we", "observe", "that", "TrufLL", "seems", "a", "promising", "approach", "to", "learn", "a", "language", "policy", "without", "a", "supervised", "training", "phase,", "by", "solely", "interacting", "with", "another", "language", "system." ]

[ 0, 13, 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 ]

[ "As", "for", "the", "scratch+KL", "samples,", "they", "are", "either", "not", "grounded,", "or", "showcase", "degenerated", "language." ]

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

[ "Despite", "a", "peaky", "distribution,", "TrufLL", "has", "moderate", "repetitions", "across", "images,", "and", "is", "mostly", "overconfident." ]

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

[ "It", "is", "remarkable", "that", "TrufLL", "with", "a", "generic", "LM", "still", "manages", "to", "capture", "the", "necessary", "subtleties", "of", "VQG,", "without", "any", "prior", "task", "knowledge." ]

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

[ "Interestingly,", "they", "are", "often", "grounded", "with", "different", "objects", "of", "the", "image." ]

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

[ "On", "the", "other", "hand,", "TrufLL", "generate", "adequate", "questions,", "resulting", "in", "the", "expected", "answer." ]

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

[ "They", "inaccurately", "capture", "the", "pragmatics", "of", "the", "task." ]

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

[ "Qualitative", "performance:", "We", "display", "qualitative", "samples", "in", "Figure", "2", "and", "Appendix", "D.", "On", "the", "one", "hand,", "the", "pretrained", "baselines", "generate", "either", "a", "question", "inconsistent", "with", "the", "visual", "context,", "or", "which", "fails", "to", "answer", "the", "expected", "answer." ]

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

[ "Less", "positively,", "TrufLL", "diversity", "metrics", "suggest", "potential", "mode", "collapse,", "with", "a", "high", "peakiness", "and", "self-BLEU", "score." ]

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

[ "Therefore,", "TrufLL", "seems", "to", "correctly", "capture", "the", "language", "distribution", "of", "the", "initial", "LM." ]

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

[ "On", "the", "other", "hand,", "TrufLL", "with", "the", "external", "LM", "reports", "poor", "dataset-based", "language", "scores,", "while", "maintaining", "a", "low", "external", "perplexity." ]

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

[ "Noticeably,", "TrufLL", "with", "the", "task-specific", "LM", "follows", "the", "same", "pattern." ]

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