sentences
sequence
ner_tags
sequence
[ "Table", "2", "shows", "that", "models", "deployed", "with", "the", "reranking", "method", "have", "relatively", "higher", "or", "com-", "parable", "Bart", "scores", ",", "but", "greatly", "improved", "perfor-", "mance", "on", "the", "F1", "score", "(", "from", ".54to.85", ")", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "I-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "I-MetricName", "O", "O", "B-MetricValue", "O", "O" ]
[ "This", "result", ",", "too", ",", "underscores", "the", "advantages", "of", "the", "reranking", "method", ".", "5.4", "Error", "Analysis", "While", "BlenderBot", "showed", "strong", "performance", "when", "using", "reranking", ",", "a", "certain", "number", "of", "gen-", "erated", "utterances", "still", "did", "not", "match", "the", "real", "tutor882", "labels", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "When", "a", "matching", "utterance", "type", "can", "not", "be", "found", "in", "a", "limited", "pool", "of", "candidates", ",", "we", "could", "have", "chosen", "to", "increase", "the", "candidate", "pool", "to", "pro-", "mote", "the", "probability", "of", "selecting", "a", "match", ".", "However", ",", "in", "this", "early", "effort", "to", "generate", "hedges", ",", "we", "want", "to", "ensure", "sufficient", "quality", "in", "the", "generated", "output", "but", "also", "explore", "the", "limitations", "of", "current", "language", "models", "for", "generating", "socially", "relevant", "phenomena", "on", "the", "basis", "of", "a", "spontaneous", "spoken", "interaction", "dataset", ".", "We", "can", "learn", "about", "the", "limitations", "of", "these", "mod-", "els", "by", "examining", "places", "where", "the", "system", "did", "not", "generate", "the", "desired", "strategy", "(", "that", "is", ",", "generated", "a", "hedge", "when", "the", "real", "tutor", "did", "not", "or", "vice", "versa", ")", ".", "We", "first", "divide", "these", "strategy", "mismatches", "into", "over-", "generation", "errors", ",", "where", "the", "generator", "generates", "a", "hedge", "where", "it", "should", "not", "and", "under", "-", "generation", "errors", "when", "it", "does", "not", "generate", "a", "hedge", "but", "should", ".", "Among", "the", "1395", "annotated", "turns", "outputted", "by", "the", "3", "generators", ",", "there", "are", "13.3", "%", "of", "over", "-", "generation", "errors", "and", "86.7", "%", "under", "-", "generation", "errors", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricValue", "I-MetricValue", "O", "B-MetricName", "I-MetricName", "I-MetricName", "I-MetricName", "O", "B-MetricValue", "I-MetricValue", "B-MetricName", "I-MetricName", "I-MetricName", "I-MetricName", "O" ]
[ "These", "errors", "are", "particularly", "interesting", "in", "the", "context", "of", "reranking", ",", "as", "it", "relied", "strongly", "on", "the", "hedge", "classi-", "fier", ".", "The", "hedge", "classifier", "selected", "the", "most", "suitable", "utterances", ",", "and", "yet", "the", "model", "still", "produced", "the", "wrong", "strategy", "-", "or", "at", "the", "very", "least", "mismatches", "with", "the", "strategy", "of", "the", "real", "tutor", ".", "Therefore", ",", "we", "analyze", "the", "generated", "utterances", "corresponding", "to", "these", "two", "types", "of", "errors", "and", "iden-", "tify", "two", "potential", "causes", ".", "First", ",", "there", "are", "still", "some", "places", "where", "the", "model", "generates", "a", "hedge", "where", "it", "should", "generate", "a", "non-", "hedge", ".", "As", "we", "mentioned", "in", "Section", "4.4", ",", "we", "invited", "humans", "to", "annotate", "the", "models", "’", "outputs", "in", "terms", "of", "hedge", "labels", ".", "We", "compare", "the", "human", "-", "annotations", "of", "the", "model", "output", "(", "where", "they", "labeled", "the", "out-", "put", "as", "hedge", "or", "non", "-", "hedge", ")", "with", "the", "output", "of", "the", "BERT", "-", "based", "classifier", "on", "the", "same", "generated", "utter-", "ances", "to", "calculate", "the", "F1", "score", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "I-MetricName", "O" ]
[ "We", "find", "that", "there", "is", "a", "difference", "of", "about", "9", "points", "between", "the", "F1", "score", "for", "human", "annotation", "(", "85", "%", ")", "shown", "in", "Table", "2", ",", "and", "the", "F1", "score", "for", "the", "same", "BERT", "-", "based", "hedge", "classifier", "(", "94", "%", ")", "reported", "in", "Raphalen", "et", "al", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricValue", "I-MetricValue", "O", "O", "B-MetricName", "I-MetricName", "O", "O", "O", "O", "B-MetricValue", "I-MetricValue", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "I-MetricName", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "O", "B-MetricValue", "I-MetricValue", "O", "O", "O", "O", "O", "O", "O" ]
[ "(", "2022", ")", ".", "We", "assume", "that", "the", "classifier", "we", "used", "may", "have", "misclassified", "some", "generated", "utterances", "and", "we", "therefore", "label", "them", "as", "Classification", "Errors", ".", "This", "category", "accounts", "for", "92.5", "%", "of", "over", "-", "generation", "er-", "rors", ",", "and", "15.3", "%", "of", "under", "-", "generation", "errors", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricValue", "I-MetricValue", "O", "B-MetricName", "I-MetricName", "I-MetricName", "I-MetricName", "I-MetricName", "O", "O", "B-MetricValue", "I-MetricValue", "O", "B-MetricName", "I-MetricName", "I-MetricName", "I-MetricName", "O" ]
[ "Second", ",", "the", "basic", "functionality", "of", "an", "end", "-", "to-", "end", "language", "model", "of", "this", "kind", "is", "to", "produce", "the", "most", "coherent", "next", "utterance", "based", "on", "the", "di", "-", "alogue", "history", ".", "This", "may", "result", "in", "the", "language", "model", "privileging", "coherence", "of", "content", "over", "style", "of", "delivery", ".", "That", "is", ",", "the", "model", "may", "not", "be", "able", "to", "find", "an", "appropriate", "strategy", "match", "among", "the", "coherent", "candidates", ",", "even", "when", "the", "candidate", "pool", "size", "is", "50", ".", "We", "label", "this", "a", "Goal", "Mismatch", "as", "the", "propositional", "or", "content", "coherence", "goals", "of", "the", "sys-", "tem", "may", "be", "trumping", "the", "social", "goals", ",", "We", "found", "84.7", "%", "in", "under", "-", "generation", "errors", "and", "7.5", "%", "in", "over-", "generation", "errors", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricValue", "I-MetricValue", "O", "B-MetricName", "I-MetricName", "I-MetricName", "I-MetricName", "O", "B-MetricValue", "I-MetricValue", "O", "B-MetricName", "I-MetricName", "I-MetricName", "O" ]
[ "18", "%", "of", "the", "cases", "where", "the", "pool", "did", "not", "include", "the", "right", "strategy", ".", "An", "example", "of", "each", "type", "of", "error", "is", "given", "in", "Fig-", "ure", "3", ".", "The", "first", "example", "belongs", "to", "the", "Classifica-", "tion", "Error", "type", ",", "where", "the", "classifier", "misclassified", "the", "system", "response", "(", "i.e", ".", "β€œ", "We", "just", "found", "that", "the", "answer", "is", "two", "x", "equals", "three", "”", ")", "as", "a", "hedge", ".", "In", "the", "second", "example", ",", "the", "tutor", "is", "trying", "to", "help", "the", "tu-", "tee", "to", "approach", "the", "answer", "step", "by", "step", ",", "but", "the", "tutee", "can", "not", "come", "up", "with", "a", "worked", "idea", ".", "Here", "it", "is", "clear", "that", "the", "tutee", "is", "flailing", "and", "it", "is", "therefore", "probably", "not", "advisable", "to", "increase", "the", "student", "’s", "stress", "with", "a", "volley", "of", "questions", "that", "the", "tutee", "can", "clearly", "not", "answer", ".", "The", "tutor", "thus", "uses", "a", "hedge", "as", "a", "response", ".", "Conversely", ",", "the", "generator", "produces", "a", "question", ".", "The", "generated", "utterance", "is", "β€œ", "What", "do", "you", "think", "we", "should", "do", ",", "what", "’s", "the", "next", "step", "”", ".", "This", "ex-", "ample", "corresponds", "to", "our", "Goal", "Mismatch", "Error", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "I-MetricName", "I-MetricName", "O" ]
[ "It", "shows", "that", "the", "generator", "may", "not", "understand", "the", "social", "context", ",", "but", "is", "looking", "for", "a", "coherent", "response", ".", "TheGoal", "Mismatch", "Error", "is", "perhaps", "the", "most", "interesting", "of", "the", "errors", ",", "and", "thus", "to", "verify", "our", "hy-", "pothesis", "β€”", "that", "the", "coherence", "goals", "of", "the", "models", "may", "impede", "the", "social", "goals", "β€”", "we", "looked", "into", "the", "nature", "of", "the", "relationship", "between", "rapport", "(", "between", "tutor", "and", "tutee", ")", "and", "the", "generation", "of", "hedges", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "I-MetricName", "I-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "As", "described", "above", ",", "Madaio", "et", "al", ".", "(", "2017", ")", "found", "that", "hedges", "are", "generated", "when", "rapport", "is", "low", ".", "Since", "our", "corpus", "contained", "rapport", "annotations", "for", "every", "30", "seconds", "of", "the", "interaction", ",", "we", "looked", "at", "the", "rap-", "port", "level", "in", "play", "when", "the", "model", "over", "-", "generated", "and", "under", "-", "generated", "hedges", ".", "Since", "rapport", "is", "anno-", "tated", "from", "1", "to", "7", "in", "the", "dataset", ",", "for", "convenience", ",", "we", "divided", "it", "into", "3", "levels", ":", "high", "(", "5", "-", "7", ")", ",", "medium", "(", "3", "-", "5", ")", ",", "and", "low", "rapport", "(", "1", "-", "3", ")", ",", "as", "shown", "in", "Table", "3", ".", "TypeRapportHigh", "Medium", "Low", "Over", "-", "generation", "0", "3", "0", "Under", "-", "generation", "13", "130", "75", "Table", "3", ":", "Goal", "Mismatch", "Errors", "Distribution883", "Tutor", ":", "mhm", ",", "so", "what", "are", "the", "two", "things", "that", "y", "ou", "can", "do", ",", "there", "are", "only", "two", ".", "Tutee", ":", "I", "could", "mo", "ve", "x", "over", "to", "three", "x", ".", "Tutor", ":", "Y", "eah", "what", "do", "y", "ou", "think", "we", "should", ",", "should", "we", "do", "it", "now", ".", "Tutee", ":", "mmm", "I", "do", "nt", "know", ",", "do", "you", "ha", "ve", "the", ",", "um", ",", "minus", "Tutor", ":", "Well", "the", "other", "one", "would", "be", "to", "divide", "three", "by", "thirty", "six", "...", "(", "hedge", ")", "System", ":", "What", "do", "y", "ou", "think", "we", "should", "do", ",", "what", "'s", "the", "next", "step", "(", "non-", "hedge", ")", "Tutor", ":", "W", "ait", ",", "you", "'re", "on", "question", "three", ",", "<", "laughter", ">", "just", "checking", "Tutee", ":", "I", "told", "y", "ou", ",", "the", "answer", "is", "two", "Tutor", ":", "W", "e", "just", "found", "out", "Tutee", ":", "oh", "really", "?", "Tutor", ":", "<", "laughter", ">", "oka", "y", "problem", "four", "is", "four", "(", "non-", "hedge", ")", "System", ":", "We", "just", "found", "out", "the", "answer", "is", "two", "x", "equals", "three", "(", "hedge", ")", "Goal", "Mismatch", "Error", "Classification", "ErrorFigure", "3", ":", "Strategy", "Mismatch", "Errors", "for", "Reranking", "Method", "As", "only", "3", "errors", "appear", "in", "the", "category", "of", "over-", "generation", "error", ",", "we", "can", "not", "obtain", "a", "meaningful", "conclusion", "due", "to", "size", ".", "However", ",", "the", "generators", "generate", "fewer", "hedges", "when", "rapport", "is", "low", ",", "an", "under", "-", "generation", "error", ",", "in", "contradiction", "to", "stud-", "ies", "showing", "that", "speakers", "are", "more", "careful", "about", "threatening", "the", "face", "of", "(", "or", "embarrassing", ")", "their", "in-", "terlocutors", "when", "the", "social", "bond", "between", "them", "is", "weak", "(", "Madaio", "et", "al", ".", ",", "2017", ")", ".", "We", "believe", "that", "this", "is", "because", "more", "hedges", "are", "found", "in", "low", "rapport", "interaction", ".", "Therefore", ",", "we", "count", "the", "hedge", "distribu-", "tion", "of", "the", "low", "and", "high", "rapport", "interaction", "in", "the", "test", "dataset", ".", "264", "hedges", "are", "found", "in", "low", "rapport", "interaction", ",", "and", "42", "in", "high", "rapport", "interaction", ".", "This", "distribution", "corresponds", "to", "the", "fact", "that", "a", "hedge", "is", "most", "likely", "to", "happen", "in", "low", "rapport", "interactions", ".", "The", "under", "-", "generation", "errors", "are", "the", "cases", "where", "there", "should", "be", "hedges", "but", "non", "-", "hedges", "were", "gener-", "ated", ".", "In", "the", "test", "dataset", ",", "more", "hedges", "occur", "in", "low", "rapport", ",", "and", "the", "generator", "under", "-", "generates", "more", "in", "low", "rapport", ",", "because", "there", "are", "more", "hedges", "that", "should", "be", "generated", "in", "low", "rapport", ".", "So", ",", "the", "gener-", "ators", "make", "more", "errors", "in", "low", "rapport", "interaction", "due", "to", "an", "imbalance", "in", "hedge", "distribution", "between", "low", "and", "high", "rapport", "interaction", ".", "Goal", "Mismatch", "error", "directly", "addresses", "our", "pri-", "mary", "question", "1", ":", "How", "effectively", "do", "end", "-", "to", "-", "end", "models", "perform", "when", "generating", "hedges", "on", "their", "own", "?" ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "I-MetricName", "I-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Due", "to", "this", "fundamental", "discrepancy", "between", "competing", "goals", ",", "end", "-", "to", "-", "end", "language", "models", "are", "unable", "to", "inherently", "learn", "and", "discern", "when", "to", "ap-", "ply", "hedges", "appropriately", ".", "5.4.1", "Lexical", "Diversity", "of", "the", "Generated", "Output", "As", "we", "have", "seen", ",", "LLMs", "can", "generate", "a", "hedge", "or", "non", "-", "hedge", "with", "the", "help", "of", "the", "reranking", "method", ".", "However", ",", "do", "language", "models", "spontaneously", "use", "different", "types", "of", "hedges", "in", "a", "human", "-", "like", "way", "?", "To", "investigate", "this", "question", ",", "we", "applied", "the", "rule", "-", "based", "hedge", "classifier", "from", "(", "Raphalen", "et", "al", ".", ",", "2022", ")", "to", "au", "-", "tomatically", "annotate", "the", "utterances", "generated", "by", "models", "in", "subcategories", "of", "hedges", "(", "as", "defined", "in", "Section", "2.1", ")", ",", "and", "we", "compare", "the", "models", "’", "and", "hu-", "mans", "’", "distributions", "of", "different", "hedge", "strategies", ".", "The", "rule", "-", "based", "classifier", "used", "linguistic", "patterns", "to", "identify", "each", "hedge", "subcategory", ".", "We", "have", "pre-", "ferred", "here", "to", "use", "the", "rule", "-", "based", "classifier", "rather", "than", "the", "machine", "learning", "classifiers", "to", "avoid", "the", "dependence", "on", "and", "bias", "of", "probabilistic", "learning-", "based", "classifiers", ".", "Indeed", ",", "learning", "-", "based", "classifiers", "may", "be", "biased", "towards", "predicting", "the", "categories", "that", "are", "the", "most", "frequent", "in", "the", "dataset", ".", "Further-", "more", ",", "the", "rule", "-", "based", "classifier", "reaches", "a", "94.7", "F1", "score", "(", "Raphalen", "et", "al", ".", ",", "2022", ")", ",", "which", "is", "compara-", "ble", "to", "the", "best", "performance", "(", "96.7", "F1", "score", ")", "using", "the", "Light", "Gradient", "-", "Boosting", "Machine", "(", "LGBM", ")", "(", "Ke", "et", "al", ".", ",", "2017", ")", "classifier", ".", "The", "above", "results", "show", "that", "the", "model", "can", "spon-", "taneously", "learn", "to", "use", "different", "types", "of", "hedges", ".", "Indeed", ",", "the", "models", "are", "capable", "of", "carrying", "out", "lin-", "guistic", "diversity", "on", "hedges", "based", "on", "learning", "from", "real", "human", "dialogues", ".", "6", "Conclusion", "and", "Future", "Work", "In", "this", "paper", ",", "we", "have", "shown", "that", "the", "reranking", "method", "helps", "LLMs", "to", "generate", "hedges", "β€”", "an", "im-", "portant", "social", "conversational", "strategy", "that", "can", "avoid", "face", "threats", "towards", "an", "interlocutor", "by", "attenuating", "an", "impact", "of", "an", "expression", ".", "We", "find", "that", "an", "implicit", "fine", "-", "tuning", "approach", "(", "i.e.", ",", "without", "any", "supervision", "by", "a", "hedge", "label", ")", "is", "not", "sufficient", "for", "generating", "hedges", ",", "but", "a", "reranking", "method", "significantly", "im-", "proves", "performance", "in", "generating", "hedges", ",", "with", "a", "final", "F1", "score", "of", ".85for", "the", "BART", "model", "and", ".84", "for", "the", "BlenderBot", "model", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "O", "O", "O", "O", "B-MetricName", "I-MetricName", "O", "B-MetricValue", "O", "B-MethodName", "O", "O", "B-MetricValue", "O", "O", "B-MethodName", "O", "O" ]
[ "We", "also", "performed", "an", "er-", "ror", "analysis", "on", "the", "generated", "results", "and", "found", "that", "two", "types", "of", "errors", "occur", "in", "the", "reranking", "method", ":", "Classification", ",", "and", "Goal", "Mismatch", ".", "The", "vast", "ma-", "jority", "of", "errors", "fall", "into", "the", "category", "of", "Goal", "Mis-", "match", ",", "indicating", "an", "important", "conflict", "between884", "contemporary", "language", "models", "’", "primary", "goal", "of", "ensuring", "coherence", "and", "the", "social", "goal", "of", "managing", "face", ",", "which", "is", "indispensable", "for", "human", "conversa-", "tion", ".", "While", "we", "were", "able", "to", "generate", "hedges", ",", "we", "were", "not", "able", "to", "necessarily", "generate", "them", "where", "they", "were", "needed", "most", ".", "That", "is", ",", "conversational", "strategies", "are", "adaptive", "in", "the", "sense", "that", "they", "re-", "spond", "to", "conversational", "strategies", "uttered", "by", "the", "previous", "speaker", "(", "Zhao", "et", "al", ".", ",", "2014", ")", ".", "We", "conclude", "that", ",", "going", "forward", ",", "we", "will", "need", "a", "way", "of", "adding", "an", "underlying", "representation", "of", "the", "social", "state", "of", "the", "dialogue", "to", "improve", "dialogue", "generation", ".", "In", "this", "paper", "we", "addressed", "the", "question", "of", "how", "to", "generate", "hedges", ",", "but", "when", "to", "generate", "hedges", "remains", "an", "important", "and", "unexplored", "question", ".", "In", "future", "work", ",", "we", "may", "first", "explore", "the", "temporal", "re-", "lationships", "between", "the", "hedge", "and", "other", "conversa-", "tional", "information", "(", "e.g.", ",", "other", "conversational", "strate-", "gies", ",", "level", "of", "rapport", ")", "by", "sequential", "rule", "mining", "techniques", ",", "then", "apply", "RL", "-", "based", "methods", "to", "inves-", "tigate", "in", "a", "more", "detailed", "manner", "the", "optimal", "way", "to", "predict", "where", "hedges", "should", "occur", ".", "In", "this", "context", ",", "we", "note", "that", "ChatGPT", "can", "generate", "a", "hedge", "when", "requested", "explicitly", "to", "do", "so", ",", "but", "does", "not", "generate", "hedges", "of", "its", "own", "volition", "(", "so", "to", "speak", ")", ",", "for", "exam-", "ple", ",", "when", "face", "-", "threatening", "acts", "such", "as", "instruction", "are", "engaged", "in", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "We", "began", "this", "paper", "by", "describing", "the", "need", "for", "hedges", "in", "instructional", "dialogues", "such", "as", "those", "engaged", "in", "by", "intelligent", "tutoring", "systems", ".", "The", "current", "dataset", "consists", "of", "authentic", "real", "-", "world", "tu-", "toring", "sessions", ",", "but", "as", "carried", "out", "by", "untrained", "teenagers", ".", "We", "note", "that", "peer", "tutoring", "is", "a", "power-", "ful", "method", "of", "teaching", ",", "used", "in", "classrooms", "around", "the", "world", ",", "and", "previous", "work", "shows", "that", "when", "un-", "trained", "peer", "tutors", "use", "hedges", ",", "their", "tutees", "attempt", "more", "problems", "and", "solve", "more", "problems", "correctly", "(", "Madaio", "et", "al", ".", ",", "2017", ")", ".", "However", ",", "they", "are", "inexperi-", "enced", "and", "so", "in", "future", "work", "it", "will", "be", "important", "to", "investigate", "the", "interaction", "between", "trained", "tutors", "and", "tutee", "as", "well", ",", "for", "instance", ",", "by", "using", "the", "Teacher-", "Student", "Chatroom", "Corpus", "(", "Caines", "et", "al", ".", ",", "2020", ")", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "I-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "We", "believe", "that", "the", "methods", "and", "results", "from", "the", "cur-", "rent", "work", "will", "facilitate", "the", "investigation", "of", "expert", "tutors", "in", "future", "research", ".", "Broader", "Impact", "Since", "the", "1990s", ",", "research", "has", "shown", "the", "the", "im-", "portance", "of", "intelligent", "tutoring", "systems", "as", "effective", "learning", "environment", ",", "s", "and", "supports", "for", "classroom", "learning", "(", "Anderson", "et", "al", ".", ",", "1995", ")", ".", "But", "virtual", "peer", "tutors", "have", "not", "yet", "achieved", "their", "potential", ",", "in", "part", "because", "of", "the", "difficulty", "of", "generating", "the", "social", "infrastructure", "of", "peer", "learning", "as", "well", "as", "the", "content", "of", "the", "matter", "being", "tutored", ".", "This", "paper", ",", "whose", "data", "comes", "from", "a", "corpus", "of", "peer", "tutoring", "dialogues", ",", "should", "therefore", "be", "seen", "as", "a", "step", "in", "the", "right", "direction", ".", "Acknowledgments", "We", "thank", "the", "anonymous", "reviewers", "for", "their", "help-", "ful", "feedback", ".", "We", "express", "sincere", "gratitude", "to", "the", "members", "of", "the", "ArticuLab", "at", "Inria", "Paris", "for", "their", "invaluable", "assistance", "in", "the", "successful", "completion", "of", "this", "research", ",", "and", "to", "the", "members", "of", "the", "Articu-", "Lab", "at", "Carnegie", "Mellon", "Pittsburgh", "for", "answering", "our", "questions", "about", "their", "prior", "work", ".", "This", "study", "re-", "ceived", "support", "from", "the", "French", "government", ",", "admin-", "istered", "by", "the", "Agence", "Nationale", "de", "la", "Recherche", ",", "as", "part", "of", "the", "\"", "Investissements", "d’avenir", "\"", "program", ",", "with", "reference", "to", "ANR-19", "-", "P3IA-0001", "(", "PRAIRIE", "3IA", "Institute", ")", ".", "Limitations", "Several", "limitations", "apply", "to", "the", "current", "study", ".", "While", "research", "shows", "that", "multimodal", "signals", "play", "an", "important", "role", "in", "conversational", "strategies", "(", "Zhao", "et", "al", ".", ",", "2016b", ")", ",", "we", "did", "not", "take", "them", "into", "account", ".", "It", "is", "an", "open", "question", "as", "to", "how", "to", "render", "large", "language", "models", "capable", "of", "generating", "multimodal", "behaviors", ".", "A", "second", "limitation", "concerns", "the", "recent", "arrival", "on", "the", "scence", "of", "ChatGPT", ",", "that", "has", "shown", "impressive", "performance", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O" ]
[ "However", "the", "models", "are", "not", "free", ",", "and", "therefore", "were", "not", "included", ".", "As", "noted", "above", ",", "another", "important", "limitation", "is", "the", "untrained", "status", "of", "the", "tutors", "in", "our", "corpus", ",", "who", "are", "teenagers", ",", "and", "not", "trained", "tutors", ".", "Their", "use", "of", "hedges", ",", "there-", "fore", ",", "comes", "from", "their", "knowledge", "of", "everyday", "so-", "cial", "interaction", ",", "and", "not", "from", "expertise", "in", "teaching", ".", "In", "looking", "at", "the", "data", ",", "we", "find", "a", "few", "places", "where", ",", "as", "instructors", "ourselves", ",", "we", "believe", "that", "a", "hedge", "is", "important", ",", "even", "though", "the", "real", "(", "teenage", ")", "tutor", "did", "not", "use", "one", ".", "The", "last", "limitation", "is", "that", ",", "while", "we", "focused", "only", "on", "generating", "hedge", "or", "non", "-", "hedge", ",", "there", "are", "ac-", "tually", "3", "different", "kinds", "of", "hedges", ",", "that", "function", "differently", ".", "We", "hope", "to", "extend", "this", "work", "and", "take885", "advantage", "of", "a", "text", "style", "transfer", "technique", "to", "gen-", "erate", "more", "kinds", "of", "hedges", "in", "future", "work", ".", "Ethical", "Statement", "The", "corpus", "used", "here", "comes", "from", "earlier", "work", "by", "the", "last", "author", "and", "her", "colleagues", ",", "and", "was", "used", "in", "accordance", "with", "the", "original", "experimenters", "’", "Insti-", "tutional", "Review", "Board", "(", "IRB", ")", ".", "Those", "experimenters", "also", "anonymised", "the", "data", ",", "removing", "any", "identify-", "ing", "information", ".", "A", "pixelated", "example", "of", "the", "video", "data", "is", "available", "at", "github.com", "/", "neuromaancer", "/", "hedge_generation", ".", "To", "counteract", "potential", "gen-", "der", "bias", "concerning", "the", "use", "of", "hedges", "in", "peer", "tu-", "toring", ",", "the", "data", "was", "collected", "from", "equal", "number", "of", "boys", "and", "girls", ".", "In", "text", "generation", "tasks", ",", "it", "is", "important", "to", "be", "aware", "of", "the", "potential", "risk", "of", "gen-", "erating", "inappropriate", "content", ".", "We", "believe", "that", ",", "in", "fact", ",", "hedges", "used", "by", "tutors", "are", "perhaps", "the", "least", "likely", "conversational", "strategy", "to", "be", "inappropriate", ",", "as", "they", "are", "the", "most", "polite", "and", "β€œ", "delicate", "”", "conver-", "sational", "moves", ".", "But", ",", "more", "generally", ",", "considerable", "additional", "work", "would", "be", "needed", "to", "filter", "out", "all", "inappropriate", "language", "for", "safe", "tutoring", "systems", "that", "engage", "in", "social", "and", "task", "interaction", ".", "References", "John", "R", "Anderson", ",", "Albert", "T", "Corbett", ",", "Kenneth", "R", "Koedinger", ",", "and", "Ray", "Pelletier", ".", "1995", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Proceedings", "of", "the", "61st", "Annual", "Meeting", "of", "the", "Association", "for", "Computational", "Linguistics", "Volume", "1", ":", "Long", "Papers", ",", "pages", "893–911", "July", "9", "-", "14", ",", "2023", "Β©", "2023", "Association", "for", "Computational", "Linguistics", "DIFFUSION", "DB", ":", "A", "Large", "-", "scale", "Prompt", "Gallery", "Dataset", "for", "Text", "-", "to", "-", "Image", "Generative", "Models", "Zijie", "J", ".", "Wang", "Evan", "Montoya", "David", "Munechika", "Haoyang", "Yang", "Benjamin", "Hoover", "Duen", "Horng", "Chau", "College", "of", "Computing", ",", "Georgia", "Tech", "{", "jayw|emontoya30|david.munechika|alexanderyang|bhoov|polo", "}", "@", "gatech.edu", "Abstract", "With", "recent", "advancements", "in", "diffusion", "models", ",", "users", "can", "generate", "high", "-", "quality", "images", "by", "writ-", "ing", "text", "prompts", "in", "natural", "language", ".", "However", ",", "generating", "images", "with", "desired", "details", "requires", "proper", "prompts", ",", "and", "it", "is", "often", "unclear", "how", "a", "model", "reacts", "to", "different", "prompts", "or", "what", "the", "best", "prompts", "are", ".", "To", "help", "researchers", "tackle", "these", "critical", "challenges", ",", "we", "introduce", "DIF-", "FUSION", "DB", ",", "the", "first", "large", "-", "scale", "text", "-", "to", "-", "image", "prompt", "dataset", "totaling", "6.5", "TB", ",", "containing", "14", "million", "images", "generated", "by", "Stable", "Diffusion", ",", "1.8", "million", "unique", "prompts", ",", "and", "hyperpa-", "rameters", "specified", "by", "real", "users", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "We", "analyze", "the", "syntactic", "and", "semantic", "characteristics", "of", "prompts", ".", "We", "pinpoint", "specific", "hyperparameter", "values", "and", "prompt", "styles", "that", "can", "lead", "to", "model", "errors", "and", "present", "evidence", "of", "potentially", "harmful", "model", "usage", ",", "such", "as", "the", "generation", "of", "misinformation", ".", "The", "unprecedented", "scale", "and", "diversity", "of", "this", "human", "-", "actuated", "dataset", "provide", "exciting", "research", "opportunities", "in", "understanding", "the", "interplay", "between", "prompts", "and", "generative", "models", ",", "detecting", "deepfakes", ",", "and", "designing", "human", "-", "AI", "interaction", "tools", "to", "help", "users", "more", "easily", "use", "these", "models", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "DIFFUSION", "DBis", "publicly", "available", "at", ":", "https", ":", "/", "/", "poloclub.github.io", "/", "diffusiondb", "." ]
[ "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "1", "Introduction", "Recent", "diffusion", "models", "have", "gained", "immense", "pop-", "ularity", "by", "enabling", "high", "-", "quality", "and", "controllable", "image", "generation", "based", "on", "text", "prompts", "written", "in", "natural", "language", "(", "Rombach", "et", "al", ".", ",", "2022", ";", "Ramesh", "et", "al", ".", ",", "2022", ";", "Saharia", "et", "al", ".", ",", "2022", ")", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Since", "the", "re-", "lease", "of", "these", "models", ",", "people", "from", "different", "do-", "mains", "have", "quickly", "applied", "them", "to", "create", "award-", "winning", "artworks", "(", "Roose", ",", "2022", ")", ",", "synthetic", "radi-", "ology", "images", "(", "Chambon", "et", "al", ".", ",", "2022", ")", ",", "and", "even", "hyper", "-", "realistic", "videos", "(", "Ho", "et", "al", ".", ",", "2022", ")", ".", "However", ",", "generating", "images", "with", "desired", "de-", "tails", "is", "difficult", ",", "as", "it", "requires", "users", "to", "write", "proper", "prompts", "specifying", "the", "exact", "expected", "results", ".", "De-", "veloping", "such", "prompts", "requires", "trial", "and", "error", ",", "Fig", ".", "1", ":", "DIFFUSION", "DBis", "the", "first", "large", "-", "scale", "dataset", "featuring", "6.5", "TB", "data", "including", "1.8", "million", "unique", "Stable", "Diffusion", "prompts", "and", "14", "million", "generated", "images", "with", "accompanying", "hyperparameters", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "It", "provides", "exciting", "research", "opportunities", "in", "prompt", "engineering", ",", "deepfake", "detection", ",", "and", "understanding", "large", "generative", "models", "." ]
[ "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "O", "B-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O" ]
[ "and", "can", "often", "feel", "random", "and", "unprincipled", "(", "Liu", "and", "Chilton", ",", "2022", ")", ".", "Willison", "et", "al", ".", "(", "2022", ")", "analo-", "gize", "writing", "prompts", "to", "wizards", "learning", "β€œ", "magical", "spells", "”", ":", "users", "do", "not", "understand", "why", "some", "prompts", "work", ",", "but", "they", "will", "add", "these", "prompts", "to", "their", "β€œ", "spell", "book", ".", "”", "For", "example", ",", "to", "generate", "highly", "-", "detailed", "im-", "ages", ",", "it", "has", "become", "a", "common", "practice", "to", "add", "spe-", "cial", "keywords", "such", "as", "β€œ", "trending", "on", "artstation", "”", "and", "β€œ", "unreal", "engine", "”", "in", "the", "prompt", ".", "Prompt", "engineering", "has", "become", "a", "field", "of", "study", "in", "the", "context", "of", "text", "-", "to", "-", "text", "generation", ",", "where", "re-", "searchers", "systematically", "investigate", "how", "to", "con-", "struct", "prompts", "to", "effectively", "solve", "different", "down-", "stream", "tasks", "(", "Branwen", ",", "2020", ";", "Reynolds", "and", "Mc-", "Donell", ",", "2021", ")", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "As", "large", "text", "-", "to", "-", "image", "models", "are", "relatively", "new", ",", "there", "is", "a", "pressing", "need", "to", "under-", "stand", "how", "these", "models", "react", "to", "prompts", ",", "how", "to", "write", "effective", "prompts", ",", "and", "how", "to", "design", "tools", "to", "help", "users", "generate", "images", "(", "Liu", "and", "Chilton", ",", "2022", ")", ".", "Our", "work", "helps", "researchers", "tackle", "these", "critical", "challenges", ",", "through", "three", "major", "contributions", ":", "β€’DIFFUSION", "DB", "(", "Fig", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O" ]
[ "1", ")", ",", "the", "first", "large", "-", "scale", "prompt", "dataset", "totaling", "6.5", "TB", ",", "containing", "14", "million", "images", "generated", "by", "Stable", "Diffu-", "sion", "(", "Rombach", "et", "al", ".", ",", "2022", ")", "using", "1.8", "million893", "Fig", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "2", ":", "DIFFUSION", "DBcontains", "14", "million", "Stable", "Diffusion", "images", ",", "1.8", "million", "unique", "text", "prompts", ",", "and", "all", "model", "hyperparameters", ":", "seed", ",", "step", ",", "CFG", "scale", ",", "sampler", ",", "and", "image", "size", "." ]
[ "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterName", "O", "B-HyperparameterName", "O", "B-HyperparameterName", "I-HyperparameterName", "O", "B-HyperparameterName", "O", "O", "B-HyperparameterName", "I-HyperparameterName", "O" ]
[ "Each", "image", "also", "has", "a", "unique", "filename", ",", "a", "hash", "of", "its", "creator", "’s", "Discord", "username", ",", "and", "a", "creation", "timestamp", ".", "To", "help", "researchers", "filter", "out", "potentially", "unsafe", "or", "harmful", "content", ",", "we", "employ", "state", "-", "of", "-", "the", "-", "art", "models", "to", "compute", "an", "NSFW", "score", "for", "each", "image", "and", "prompt", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "I-MetricName", "O", "O", "O", "O", "O", "O" ]
[ "unique", "prompts", "and", "hyperparameters", "specified", "by", "real", "users", ".", "We", "construct", "this", "dataset", "by", "collect-", "ing", "images", "shared", "on", "the", "Stable", "Diffusion", "public", "Discord", "server", "(", "Β§", "2", ")", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "We", "release", "DIFFUSION", "DB", "with", "a", "CC0", "1.0", "license", ",", "allowing", "users", "to", "flexi-", "bly", "share", "and", "adapt", "the", "dataset", "for", "their", "use", "." ]
[ "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "In", "addition", ",", "we", "open", "-", "source", "our", "code1that", "collects", ",", "processes", ",", "and", "analyzes", "the", "images", "and", "prompts", ".", "β€’Revealing", "prompt", "patterns", "and", "model", "errors", ".", "The", "unprecedented", "scale", "of", "DIFFUSION", "DB", "paves", "the", "path", "for", "researchers", "to", "systematically", "investigate", "diverse", "prompts", "and", "associated", "im-", "ages", "that", "were", "previously", "not", "possible", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "By", "char-", "acterizing", "prompts", "and", "images", ",", "we", "discover", "com-", "mon", "prompt", "patterns", "and", "find", "different", "distribu-", "tions", "of", "the", "semantic", "representations", "of", "prompts", "and", "images", ".", "Our", "error", "analysis", "highlights", "partic-", "ular", "hyperparameters", "and", "prompt", "styles", "can", "lead", "to", "model", "errors", ".", "Finally", ",", "we", "provide", "evidence", "of", "image", "generative", "models", "being", "used", "for", "poten-", "tially", "harmful", "purposes", "such", "as", "generating", "misin-", "formation", "and", "nonconsensual", "pornography", "(", "Β§", "3", ")", ".", "β€’Highlighting", "new", "research", "directions", ".", "As", "the", "first", "-", "of", "-", "its", "-", "kind", "text", "-", "to", "-", "image", "prompt", "dataset", ",", "DIFFUSION", "DBopens", "up", "unique", "opportunities", "for", "researchers", "from", "natural", "language", "processing", "(", "NLP", ")", ",", "computer", "vision", ",", "and", "human", "-", "computer", "interaction", "(", "HCI", ")", "communities", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "The", "scale", "and", "diversity", "of", "this", "human", "-", "actuated", "dataset", "will", "provide", "new", "research", "opportunities", "in", "better", "tooling", "for", "prompt", "engineering", ",", "explaining", "large", "generative", "models", ",", "and", "detecting", "deepfakes", "(", "Β§", "4", ")", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "O", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "O", "O", "B-TaskName", "I-TaskName", "O", "O", "O", "O", "O" ]
[ "We", "believe", "DIFFUSION", "DBwill", "serve", "as", "an", "im-", "portant", "resource", "for", "researchers", "to", "study", "the", "roles", "of", "prompts", "in", "text", "-", "to", "-", "image", "generation", "and", "design", "next", "-", "generation", "human", "-", "AI", "interaction", "tools", "." ]
[ "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "1Code", ":", "https", ":", "/", "/", "github.com", "/", "poloclub", "/", "diffusiondb2", "Constructing", "D", "IFFUSION", "DB", "We", "construct", "DIFFUSION", "DB", "(", "Fig", ".", "2", ")", "by", "scraping", "user", "-", "generated", "images", "from", "the", "official", "Stable", "Dif-", "fusion", "Discord", "server", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O" ]
[ "We", "choose", "Stable", "Diffusion", "as", "it", "is", "currently", "the", "only", "open", "-", "source", "large", "text", "-", "to-", "image", "generative", "model", ",", "and", "all", "generated", "images", "have", "a", "CC0", "1.0", "license", "that", "allows", "uses", "for", "any", "pur-", "pose", "(", "StabilityAI", ",", "2022b", ")", "." ]
[ "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "We", "choose", "the", "official", "public", "Discord", "server", "as", "it", "has", "strict", "rules", "against", "generating", "illegal", ",", "hateful", ",", "or", "NSFW", "(", "not", "suitable", "for", "work", ",", "such", "as", "sexual", "and", "violent", "content", ")", "im-", "ages", ",", "and", "it", "prohibits", "sharing", "prompts", "with", "personal", "information", "(", "StabilityAI", ",", "2022a", ")", ".", "Our", "construction", "process", "includes", "collecting", "im-", "ages", "(", "Β§", "2.1", ")", ",", "linking", "them", "to", "prompts", "and", "hyperpa-", "rameters", "(", "Β§", "2.2", ")", ",", "applying", "NSFW", "detectors", "(", "Β§", "2.3", ")", ",", "creating", "a", "flexible", "file", "structure", "(", "Β§", "2.4", ")", ",", "and", "dis-", "tributing", "the", "dataset", "(", "Β§", "2.5", ")", ".", "We", "discuss", "DIFFU", "-", "SION", "DB", "’s", "limitations", "and", "broader", "impacts", "in", "Β§", "7", ",", "Β§", "8", ",", "and", "a", "Data", "Sheet", "(", "Gebru", "et", "al", ".", ",", "2020", ")", "(", "‑", "A", ")", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "I-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "2.1", "Collecting", "User", "Generated", "Images", "We", "download", "chat", "messages", "from", "the", "Stable", "Diffusion", "Discord", "channels", "with", "DiscordChatEx-", "porter", "(", "Holub", ",", "2017", ")", ",", "saving", "them", "as", "HTML", "files", ".", "We", "focus", "on", "channels", "where", "users", "can", "command", "a", "bot", "to", "run", "Stable", "Diffusion", "Version", "1", "to", "generate", "images", "by", "typing", "a", "prompt", ",", "hyperparameters", ",", "and", "the", "number", "of", "images", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "The", "bot", "then", "replies", "with", "the", "generated", "images", "and", "used", "random", "seeds", ".", "2.2", "Extracting", "Image", "Metadata", "We", "use", "Beautiful", "Soup", "(", "Richardson", ",", "2007", ")", "to", "parse", "HTML", "files", ",", "mapping", "generated", "images", "with", "their", "prompts", ",", "hyperparameters", ",", "seeds", ",", "timestamps", ",", "and", "the", "requester", "’s", "Discord", "usernames", ".", "Some", "images", "are", "collages", ",", "where", "the", "bot", "combines", "ngenerated894", "images", "as", "a", "grid", "(", "e.g.", ",", "a", "3Γ—3grid", "of", "n=", "9images", ")", ";", "these", "images", "have", "the", "same", "prompt", "and", "hyperpa-", "rameters", "but", "different", "seeds", ".", "We", "use", "Pillow", "(", "Clark", ",", "2015", ")", "to", "split", "a", "collage", "into", "nindividual", "images", "and", "assign", "them", "with", "the", "correct", "metadata", "and", "unique", "filenames", ".", "Finally", ",", "we", "compress", "all", "images", "in", "DIF-", "FUSION", "DB", "using", "lossless", "WebP", "(", "Google", ",", "2010", ")", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "2.3", "Identifying", "NSFW", "Content", "The", "Stable", "Diffusion", "Discord", "server", "prohibits", "gen-", "erating", "NSFW", "images", "(", "StabilityAI", ",", "2022a", ")", ".", "Also", ",", "Stable", "Diffusion", "has", "a", "built", "-", "in", "NSFW", "filter", "that", "automatically", "blurs", "generated", "images", "if", "it", "detects", "NSFW", "content", ".", "However", ",", "we", "find", "DIFFUSION", "DB", "still", "includes", "NSFW", "images", "that", "were", "not", "detected", "by", "the", "built", "-", "in", "filter", "or", "removed", "by", "server", "moder-", "ators", ".", "To", "help", "researchers", "filter", "these", "images", ",", "we", "apply", "state", "-", "of", "-", "the", "-", "art", "NSFW", "classifiers", "to", "compute", "NSFW", "scores", "for", "each", "prompt", "and", "image", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "I-MetricName", "O", "O", "O", "O", "O", "O" ]
[ "Re-", "searchers", "can", "determine", "a", "suitable", "threshold", "to", "fil-", "ter", "out", "potentially", "unsafe", "data", "for", "their", "tasks", ".", "NSFW", "Prompts", ".", "We", "use", "a", "pre", "-", "trained", "multilin-", "gual", "toxicity", "prediction", "model", "to", "detect", "unsafe", "prompts", "(", "Hanu", "and", "Unitary", "team", ",", "2020", ")", ".", "This", "model", "outputs", "the", "probabilities", "of", "a", "sentence", "be-", "ing", "toxic", ",", "obscene", ",", "threat", ",", "insult", ",", "identity", "attack", ",", "and", "sexually", "explicit", ".", "We", "compute", "the", "text", "NSFW", "score", "by", "taking", "the", "maximum", "of", "the", "probabilities", "of", "being", "toxic", "and", "sexually", "explicit", "(", "Fig", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "I-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "3", "Top", ")", ".", "NSFW", "Images", ".", "We", "use", "a", "pre", "-", "trained", "Efficient-", "Net", "classifier", "to", "detect", "images", "with", "sexual", "con-", "tent", "(", "Schuhmann", "et", "al", ".", ",", "2022", ")", ".", "This", "model", "predicts", "the", "probabilities", "of", "five", "image", "types", ":", "drawing", ",", "hen-", "tai", ",", "neutral", ",", "sexual", ",", "or", "porn", ".", "We", "compute", "the", "image", "NSFW", "score", "by", "summing", "the", "probabilities", "of", "hen-", "tai", ",", "sexual", ",", "and", "porn", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "I-MetricName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "We", "use", "a", "Laplacian", "convolu-", "tion", "kernel", "with", "a", "threshold", "of", "10to", "detect", "images", "that", "have", "already", "been", "blurred", "by", "Stable", "Diffusion", "and", "assign", "them", "a", "score", "of", "2.0", "(", "Fig", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "3", "Bottom", ")", ".", "As", "Stable", "Diffusion", "’s", "blur", "effect", "is", "strong", ",", "our", "blurred", "image", "detector", "has", "high", "precision", "and", "recall", "(", "both", "100", "%", "on", "50k", "randomly", "sampled", "images", ")", "." ]
[ "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "NSFW", "Detector", "Accuracy", ".", "To", "access", "the", "accu-", "racy", "of", "these", "two", "pre", "-", "trained", "state", "-", "of", "-", "the", "-", "art", "NSFW", "detectors", ",", "we", "randomly", "sample", "5k", "images", "and", "2k", "prompt", "texts", "and", "manually", "annotate", "them", "with", "two", "binary", "NSFW", "labels", "(", "one", "for", "image", "and", "one", "for", "prompt", ")", "and", "analyze", "the", "results", ".", "As", "the", "percent-", "age", "of", "samples", "predicted", "as", "NSFW", "(", "score", ">", "0.5", ")", "is", "small", ",", "we", "up", "-", "sample", "positive", "samples", "for", "annota-", "Fig", ".", "3", ":", "To", "help", "researchers", "filter", "out", "potentially", "unsafe", "data", "in", "DIFFUSION", "DB", ",", "we", "apply", "NSFW", "detectors", "to", "predict", "the", "probability", "that", "an", "image", "-", "prompt", "pair", "con-", "tains", "NSFW", "content", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "For", "images", ",", "a", "score", "of", "2.0indi-", "cates", "the", "image", "has", "been", "blurred", "by", "Stable", "Diffusion", ".", "tion", ",", "where", "we", "have", "an", "equal", "number", "of", "positive", "and", "negative", "examples", "in", "our", "annotation", "sample", ".", "After", "annotation", ",", "we", "compute", "the", "precisions", "and", "recalls", ".", "Because", "we", "have", "up", "-", "sampled", "positive", "pre-", "dictions", ",", "we", "adjust", "the", "recalls", "by", "multiplying", "false", "negatives", "by", "a", "scalar", "to", "adjust", "the", "sampling", "bias", ".", "The", "up", "-", "sampling", "does", "not", "affect", "precisions", ".", "Fi-", "nally", ",", "the", "precisions", ",", "recalls", "and", "adjusted", "recalls", "are0.3604", ",", "0.9565", ",", "and", "0.6661", "for", "the", "prompt", "NSFW", "detector", ",", "and", "0.315", ",", "0.9722", ",", "and", "0.3037", "for", "the", "image", "NSFW", "detector", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "O", "B-MetricName", "O", "B-MetricName", "I-MetricName", "B-MetricValue", "O", "B-MetricValue", "O", "O", "B-MetricValue", "O", "O", "O", "O", "O", "O", "O", "B-MetricValue", "O", "B-MetricValue", "O", "O", "B-MetricValue", "O", "O", "O", "O", "O", "O" ]
[ "Our", "results", "suggest", "two", "detectors", "are", "progressive", "classifiers", ".", "The", "lower", "adjusted", "recall", "of", "the", "prompt", "NSFW", "detector", "can", "be", "attributed", "to", "several", "potential", "factors", ",", "including", "the", "use", "of", "a", "fixed", "binary", "threshold", "and", "the", "poten-", "tial", "discrepancy", "in", "the", "definition", "of", "NSFW", "prompts", "between", "the", "detector", "and", "our", "annotation", "process", ".", "2.4", "Organizing", "D", "IFFUSION", "DB", "We", "organize", "DIFFUSION", "DBusing", "a", "flexible", "file", "structure", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "I-DatasetName", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O" ]
[ "We", "first", "give", "each", "image", "a", "unique", "file-", "name", "using", "Universally", "Unique", "Identifier", "(", "UUID", ",", "Version", "4", ")", "(", "Leach", "et", "al", ".", ",", "2005", ")", ".", "Then", ",", "we", "or-", "ganize", "images", "into", "14,000", "sub", "-", "folders", "β€”", "each", "in-", "cludes", "1,000", "images", ".", "Each", "sub", "-", "folder", "also", "includes", "a", "JSON", "file", "that", "contains", "1,000", "key", "-", "value", "pairs", "mapping", "an", "image", "name", "to", "its", "metadata", ".", "An", "exam-", "ple", "of", "this", "image", "-", "prompt", "pair", "can", "be", "seen", "in", "Fig", ".", "2", ".", "This", "modular", "file", "structure", "enables", "researchers", "to", "flexibly", "use", "a", "subset", "of", "D", "IFFUSION", "DB", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "I-DatasetName", "O" ]
[ "We", "create", "a", "metadata", "table", "in", "Apache", "Parquet", "format", "(", "Apache", ",", "2013", ")", "with", "13", "columns", ":", "unique", "image", "name", ",", "image", "path", ",", "prompt", ",", "seed", ",", "CFG", "scale", ",", "sampler", ",", "width", ",", "height", ",", "username", "hash", ",", "timestamp", ",", "image", "NSFW", "score", ",", "and", "prompt", "NSFW895", "Fig", ".", "4", ":", "The", "distribution", "of", "token", "counts", "for", "all", "1.8", "mil-", "lion", "unique", "prompts", "in", "DIFFUSION", "DB", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O" ]
[ "It", "is", "worth", "not-", "ing", "that", "Stable", "Diffusion", "truncates", "prompts", "at", "75", "tokens", "." ]
[ "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O" ]
[ "score", ".", "We", "store", "the", "table", "in", "a", "column", "-", "based", "format", "for", "efficient", "querying", "of", "individual", "columns", ".", "2.5", "Distributing", "D", "IFFUSION", "DB", "We", "distribute", "DIFFUSION", "DBby", "bundling", "each", "im-", "age", "sub", "-", "folder", "as", "a", "Zip", "file", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "We", "collect", "Discord", "usernames", "of", "image", "creators", "(", "Β§", "2.2", ")", ",", "but", "only", "in-", "clude", "their", "SHA256", "hashes", "in", "the", "distribution", "β€”", "as", "some", "prompts", "may", "include", "sensitive", "information", ",", "and", "explicitly", "linking", "them", "to", "their", "creators", "can", "cause", "harm", ".", "We", "host", "our", "dataset", "on", "a", "publicly", "ac-", "cessible", "repository2under", "a", "CC0", "1.0", "license", ".", "We", "provide", "scripts", "that", "allow", "users", "to", "download", "and", "load", "DIFFUSION", "DBby", "writing", "two", "lines", "of", "code", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O" ]
[ "We", "discuss", "the", "broader", "impacts", "of", "our", "distribution", "in", "Β§", "7", ",", "Β§", "8", ",", "and", "the", "Data", "Sheet", "(", "‑A", ")", ".", "To", "mitigate", "the", "potential", "harms", ",", "we", "provide", "a", "form", "for", "people", "to", "report", "harmful", "content", "for", "removal", ".", "Image", "cre-", "ators", "can", "also", "use", "this", "form", "to", "remove", "their", "images", ".", "3", "Data", "Analysis", "To", "gain", "a", "comprehensive", "understanding", "of", "the", "dataset", ",", "we", "analyze", "it", "from", "different", "perspectives", ".", "We", "examine", "prompt", "length", "(", "Β§", "3.1", ")", ",", "language", "(", "Β§", "3.2", ")", ",", "characteristics", "of", "both", "prompts", "(", "Β§", "3.3", ")", "and", "im-", "ages", "(", "Β§", "3.4", ")", ".", "We", "conduct", "an", "error", "analysis", "on", "misaligned", "prompt", "-", "image", "pairs", "(", "Β§", "3.5", ")", "and", "provide", "empirical", "evidence", "of", "potentially", "harmful", "uses", "of", "image", "generative", "models", "(", "Β§", "3.6", ")", ".", "3.1", "Prompt", "Length", "We", "collect", "prompts", "from", "Discord", ",", "where", "users", "can", "submit", "one", "prompt", "to", "generate", "multiple", "images", "and", "experiment", "with", "different", "hyperparameters", ".", "Our", "dataset", "contains", "1,819,808unique", "prompts", ".", "We", "tokenize", "prompts", "using", "the", "same", "tokenizer", "as", "used", "in", "Stable", "Diffusion", "(", "Platen", "et", "al", ".", ",", "2022", ")", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "This", "tokenizer", "truncates", "tokenized", "prompts", "at", "75to-", "kens", ",", "excluding", "special", "tokens", "<", "|startoftext|", ">", "2Public", "dataset", "repository", ":", "https", ":", "/", "/", "huggingface.co", "/", "datasets", "/", "poloclub", "/", "diffusiondband", "<", "|endoftext|", ">", ".", "We", "measure", "the", "length", "of", "prompts", "by", "their", "tokenized", "length", ".", "The", "prompt", "length", "distribution", "(", "Fig", ".", "4", ")", "indicates", "that", "shorter", "prompts", "(", "e.g.", ",", "around", "6", "to", "12", "tokens", ")", "are", "the", "most", "popular", ".", "The", "spike", "at", "75", "suggests", "many", "users", "sub-", "mitted", "prompts", "longer", "than", "the", "model", "’s", "limit", ",", "high-", "lighting", "the", "need", "for", "user", "interfaces", "guiding", "users", "to", "write", "prompts", "within", "the", "token", "limit", ".", "3.2", "Prompt", "Language", "We", "use", "a", "pre", "-", "trained", "language", "detector", "(", "Joulin", "et", "al", ".", ",", "2017", ")", "to", "identify", "the", "languages", "used", "in", "prompts", ".", "98.3", "%", "of", "the", "unique", "prompts", "in", "our", "dataset", "are", "written", "in", "English", ".", "However", ",", "we", "also", "find", "a", "large", "number", "of", "non", "-", "English", "languages", ",", "with", "the", "top", "four", "being", "German", "(", "5.2k", "unique", "prompts", ")", ",", "French", "(", "4.6k", ")", ",", "Italian", "(", "3.2k", ")", ",", "and", "Spanish", "(", "3k", ")", ".", "The", "lan-", "guage", "detector", "identifies", "34", "languages", "with", "at", "least", "100", "unique", "prompts", "in", "total", ".", "Stable", "Diffusion", "is", "trained", "on", "LAION-2B", "(", "en", ")", "(", "Schuhmann", "et", "al", ".", ",", "2022", ")", "that", "primarily", "includes", "images", "with", "English", "de-", "scriptions", ",", "thus", "our", "findings", "suggest", "that", "expanding", "the", "training", "data", "’s", "language", "coverage", "to", "improve", "the", "user", "experience", "for", "non", "-", "English", "communities", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "B-DatasetName", "I-DatasetName", "I-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "3.3", "Characterizing", "Prompts", "In", "this", "section", ",", "we", "explore", "the", "characteristics", "of", "prompts", "in", "DIFFUSION", "DB", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O" ]
[ "We", "examine", "the", "syn-", "tactic", "(", "Β§", "3.3.1", ")", "and", "semantic", "(", "Β§", "3.3.2", ")", "features", "of", "prompt", "text", "via", "interactive", "data", "visualizations", ".", "Lastly", ",", "We", "discuss", "the", "implications", "of", "our", "findings", "and", "suggest", "future", "research", "directions", ".", "3.3.1", "Prompt", "Syntactic", "Features", "To", "characterize", "the", "composition", "of", "prompts", ",", "we", "parse", "phrases", "from", "all", "1.8", "M", "unique", "prompts", ".", "We", "split", "each", "prompt", "by", "commas", "and", "then", "extract", "named", "entities", "(", "NE", ")", "and", "noun", "phrases", "(", "NP", ")", "from", "each", "separated", "component", "using", "use", "Spacy", "(", "Hon-", "nibal", "et", "al", ".", ",", "2020", ")", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "If", "there", "is", "no", "noun", "phrase", "in", "a", "comma", "-", "separated", "component", ",", "we", "extract", "the", "whole", "component", "(", "C", ")", "as", "a", "phrase", ".", "We", "keep", "track", "of", "each", "NP", "’s", "root", "to", "create", "a", "hierarchy", "of", "noun", "phrases", ".", "For", "example", ",", "for", "the", "prompt", "β€œ", "draw", "baby", "yoda", "in", "a", "loading", "screen", "for", "grand", "theft", "auto", "5", ",", "highly", "detailed", ",", "digital", "art", ",", "concept", "art", ",", "”", "we", "extract", "six", "phrases", ":", "β€œ", "baby", "yoda", "”", "(", "NE", ")", ",", "β€œ", "a", "loading", "screen", "”", "(", "NP", "with", "root", "β€œ", "screen", "”", ")", ",", "β€œ", "grand", "theft", "auto", "5", "”", "(", "NE", ")", ",", "β€œ", "highly", "detailed", "”", "(", "C", ")", ",", "β€œ", "digital", "art", "’", "(", "NP", "with", "root", "β€œ", "art", "”", ")", ",", "and", "β€œ", "concept", "art", "”", "(", "NP", "with", "root", "β€œ", "art", "”", ")", ".", "We", "group896", "Fig", ".", "5", ":", "We", "identify", "and", "group", "popular", "phrases", "in", "prompts", "through", "named", "entity", "recognition", "and", "dependency", "parsing", ".", "Our", "interactive", "circle", "-", "packing", "visualization", "highlights", "the", "distribution", "and", "hierarchy", "of", "these", "phrases", ".", "(", "A", ")", "The", "Overview", "visualizes", "each", "phrase", "as", "a", "circle", ",", "with", "its", "size", "representing", "the", "phrase", "’s", "frequency", ".", "In", "this", "example", ",", "a", "viewer", "clicks", "a", "circle", "to", "zoom", "into", "the", "β€œ", "painting", "”", "phrase", ".", "(", "B1", ")", "TheDetail", "View", "shows", "all", "noun", "phrases", "that", "use", "β€œ", "painting", "”", "as", "their", "root", ".", "(", "B2", ")", "Similarly", ",", "it", "shows", "all", "phrases", "that", "include", "β€œ", "oil", "painting", "”", "when", "the", "viewer", "zooms", "into", "β€œ", "oil", "painting", ".", "”", "β€œ", "digital", "art", "”", "and", "β€œ", "concept", "art", "”", "into", "the", "same", "hierarchy", "as", "they", "share", "the", "same", "NP", "root", "β€œ", "art", ".", "”", "Visualizing", "Prompt", "Phrases", ".", "We", "create", "an", "in-", "teractive", "circle", "packing", "visualization3to", "gain", "an", "understanding", "of", "the", "distribution", "and", "relationships", "between", "different", "phrases", "(", "Fig", ".", "5", ")", ".", "Circle", "pack-", "ing", "(", "Wang", "et", "al", ".", ",", "2006", ")", "is", "a", "technique", "to", "visualize", "hierarchical", "data", ",", "and", "each", "phrase", "is", "represented", "as", "a", "circle", "whose", "size", "encodes", "the", "phrase", "’s", "frequency", "in", "the", "dataset", ".", "We", "position", "sibling", "noun", "phrases", "(", "e.g.", ",", "phrases", "sharing", "the", "same", "NP", "root", ")", "inside", "their", "parent", "phrase", "’s", "circle", "through", "a", "front", "-", "chain", "packing", "algorithm", "(", "Wang", "et", "al", ".", ",", "2006", ")", ".", "Viewers", "can", "hover", "over", "a", "circle", "to", "see", "the", "corresponding", "phrase", "and", "its", "frequency", ".", "Viewers", "can", "also", "click", "a", "circle", "(", "Fig", ".", "5A", ")", "to", "zoom", "into", "that", "sub", "-", "tree", "to", "see", "more", "details", "about", "a", "phrase", "(", "Fig", ".", "5", "-", "B1", ")", "or", "a", "sub", "-", "phrase", "(", "Fig", ".", "5", "-", "B2", ")", ".", "Insights", "and", "implications", ".", "Our", "interactive", "visu-", "alization", "reveals", "that", "key", "phrases", "such", "as", "β€œ", "highly", "detailed", ",", "”", "β€œ", "intricate", ",", "”", "and", "β€œ", "greg", "rutkowski", "”", "3Phrase", "visualization", ":", "https", ":", "/", "/", "poloclub.github.io", "/", "diffusiondb", "/", "explorer", "#", "phraseare", "commonly", "used", "in", "prompts", "(", "Fig", ".", "5A", ")", ".", "The", "hierarchical", "visualization", "also", "surfaces", "popular", "image", "styles", "specified", "by", "users", ",", "such", "β€œ", "digital", "painting", ",", "”", "β€œ", "oil", "painting", ",", "”", "and", "β€œ", "portrait", "painting", "”", "for", "painting", "styles", "(", "Fig", ".", "5", "-", "B1", ")", "and", "β€œ", "studio", "lighting", ",", "”", "β€œ", "volumetric", "lighting", "”", ",", "and", "β€œ", "atmospheric", "lighting", "”", "for", "lighting", ".", "These", "phrases", "can", "be", "unfamiliar", "to", "Stable", "Diffusion", "users", ",", "especially", "beginners", ",", "which", "highlights", "the", "impor-", "tance", "of", "helping", "users", "develop", "prompting", "vocab-", "ularies", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Researchers", "can", "leverage", "DIFFUSION", "DB", "and", "our", "visualization", "to", "design", "tutorials", "and", "user", "interfaces", "that", "integrate", "exemplar", "prompts", "to", "guide", "users", "in", "describing", "their", "desired", "images", "." ]
[ "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "3.3.2", "Prompt", "Semantic", "Features", "In", "addition", "to", "analyzing", "the", "syntactic", "characteris-", "tics", "of", "prompts", ",", "we", "also", "analyze", "their", "semantic", "fea-", "tures", ".", "We", "use", "a", "pre", "-", "trained", "CLIP", "model", "(", "Radford", "et", "al", ".", ",", "2021", ")", "to", "extract", "semantic", "features", "(", "Ramesh", "et", "al", ".", ",", "2022", ")", ".", "We", "use", "a", "frozen", "CLIP", "ViT", "-", "L", "/", "14", "text", "encoder", "(", "the", "same", "model", "used", "in", "Stable", "Diffusion", ")", "to", "convert", "prompts", "into", "768", "-", "dimension", "vectors.897", "Fig", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "6", ":", "An", "interactive", "contour", "plot", "of", "1.8", "M", "prompts", "’", "CLIP", "embeddings", ",", "created", "with", "UMAP", "and", "kernel", "den-", "sity", "estimation", ".", "Text", "labels", "show", "the", "top", "keywords", "of", "prompts", "in", "a", "grid", "tile", ".", "It", "reveals", "popular", "prompt", "topics", ".", "Visualizing", "Prompt", "Embeddings", ".", "To", "study", "the", "distribution", "of", "prompts", "in", "high", "-", "dimensional", "space", ",", "we", "use", "UMAP", "(", "McInnes", "et", "al", ".", ",", "2020", ")", "to", "project", "768", "-", "dimensional", "vectors", "into", "2", "-", "D", "vectors", "for", "easy", "visualization", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "UMAP", "is", "a", "popular", "dimensional-", "ity", "reduction", "technique", "that", "is", "better", "at", "preserv-", "ing", "the", "global", "structure", "of", "data", "and", "more", "scal-", "able", "to", "large", "datasets", "compared", "to", "t", "-", "SNE", "(", "van", "der", "Maaten", "and", "Hinton", ",", "2008", ")", "and", "PCA", "(", "Hotelling", ",", "1936", ")", "." ]
[ "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O" ]
[ "We", "use", "grid", "search", "to", "fine", "-", "tune", "hyperpa-", "rameters", "n_neighbors", "(", "60", ")", "and", "min_dist", "(", "0.1", ")", "so", "that", "prompts", "are", "more", "spread", "out", "in", "a", "2", "-", "D", "space", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterName", "O", "B-HyperparameterValue", "O", "O", "B-HyperparameterName", "O", "B-HyperparameterValue", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "We", "develop", "an", "interactive", "visualization", "tool4to", "explore", "prompts", "’", "semantic", "embeddings", "(", "Fig", ".", "6", ")", ".", "We", "use", "Kernel", "Density", "Estimation", "(", "KDE", ")", "(", "Rosen-", "blatt", ",", "1956", ")", "with", "a", "standard", "multivariate", "Gaussian", "kernel", "and", "Silverman", "bandwidth", "(", "Silverman", ",", "2018", ")", "to", "estimate", "the", "distribution", "of", "prompts", "’", "UMAP", "rep-", "resentations", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "O", "O", "O" ]
[ "Then", ",", "we", "visualize", "the", "estimated", "dis-", "tribution", "as", "a", "contour", "plot", ".", "To", "summarize", "prompts", "that", "are", "in", "the", "same", "region", ",", "we", "create", "four", "grids", "with", "varying", "granularity", "and", "pre", "-", "compute", "key-", "words", "for", "each", "grid", "tile", ",", "by", "treating", "all", "prompts", "in", "the", "tile", "as", "a", "document", "and", "selecting", "the", "top", "4", "keywords", "with", "the", "highest", "TF", "-", "IDF", "scores", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "I-MetricName", "I-MetricName", "I-MetricName", "O" ]
[ "Interactions", ".", "Our", "visualization", "shows", "keywords", "of", "tiles", "that", "are", "close", "to", "high", "-", "density", "regions", "and", "prompt", "clusters", "by", "default", ".", "Viewers", "can", "hover", "over", "a", "tile", "to", "see", "its", "keywords", ",", "pan", "and", "zoom", "in", "to", "see", "more", "details", "of", "specific", "regions", ",", "and", "click", "a", "button", "to", "display", "each", "prompt", "as", "a", "small", "dot", "that", "viewers", "can", "hover", "over", "to", "read", "its", "prompt", "text", ".", "4Prompt", "embedding", "visualization", ":", "https", ":", "/", "/", "poloclub", ".", "github.io", "/", "diffusiondb", "/", "explorer", "/", "#", "prompt-embedding", "Fig", ".", "7", ":", "CLIP", "embeddings", "of", "2", "M", "randomly", "selected", "images", "shown", "as", "a", "contour", "plot", ",", "with", "text", "labels", "being", "keywords", "of", "prompts", "in", "the", "grid", "tiles", ".", "It", "shows", "images", "have", "a", "different", "embedding", "distribution", "from", "prompts", ".", "Insights", "and", "implications", ".", "Our", "semantic", "embed-", "ding", "visualization", "(", "Fig", ".", "6", ")", "highlights", "two", "popular", "prompt", "categories", ":", "art", "-", "related", "prompts", "(", "left", "in", "the", "plot", ")", "and", "photography", "-", "related", "prompts", "(", "dark", "blue", "regions", "on", "the", "right", ")", ".", "These", "two", "groups", "appear", "distant", "from", "each", "other", "in", "the", "UMAP", "space", ",", "sug-", "gesting", "that", "the", "prompts", "for", "art", "and", "photography", "typically", "have", "distinct", "semantic", "representations", ".", "In-", "terestingly", ",", "photography", "prompts", "appear", "to", "contain", "two", "clusters", ":", "one", "for", "non", "-", "human", "objects", "(", "top", "right", ")", "and", "another", "for", "celebrities", "(", "bottom", "right", ")", ".", "Small", "prompt", "clusters", "outside", "the", "central", "area", "often", "fea-", "ture", "artist", "names", ".", "Our", "findings", "suggest", "that", "future", "researchers", "can", "leverage", "the", "prompt", "usage", "distri-", "bution", "to", "fine", "-", "tune", "generative", "models", "to", "tailor", "to", "specific", "popular", "prompt", "categories", ".", "3.4", "Characterizing", "Images", "We", "visualize5the", "CLIP", "embedding", "distribution", "of", "2", "million", "unique", "image", "instances", "randomly", "sam-", "pled", "from", "DIFFUSION", "DB", "(", "Fig", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O" ]
[ "7", ")", "by", "defining", "the", "unique", "key", "as", "the", "combination", "of", "the", "image", "’s", "prompt", "and", "hyperparameters", "CFG", "scale", ",", "step", ",", "size", ",", "and", "seed", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterName", "I-HyperparameterName", "O", "B-HyperparameterName", "O", "B-HyperparameterName", "O", "O", "B-HyperparameterName", "O" ]
[ "We", "use", "the", "UMAP", "model", "that", "was", "previously", "trained", "on", "the", "prompt", "embeddings", "to", "project", "the", "image", "embeddings", "into", "the", "same", "2", "-", "D", "space", "." ]
[ "O", "O", "O", "B-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Finally", ",", "we", "apply", "the", "same", "method", "we", "used", "for", "our", "prompt", "embedding", "visualization", "(", "Β§", "3.3.2", ")", "to", "generate", "a", "contour", "plot", "and", "grid", "label", "overlays", ".", "Insights", "and", "implications", ".", "Our", "image", "embed-", "ding", "visualization", "reveals", "that", "generated", "images", "have", "a", "different", "distribution", "from", "their", "prompts", "in", "the", "CLIP", "embedding", "space", ".", "For", "example", ",", "the", "5Image", "embedding", "visualization", ":", "https", ":", "/", "/", "poloclub", ".", "github.io", "/", "diffusiondb", "/", "explorer", "/", "#", "image-embedding898", "β€œ", "movie", "”", "cluster", "in", "the", "prompt", "embedding", "has", "been", "replaced", "by", "the", "β€œ", "portrait", "”", "cluster", "in", "the", "image", "em-", "bedding", ".", "This", "suggests", "the", "semantic", "representa-", "tions", "of", "prompts", "and", "their", "generated", "images", "may", "not", "be", "perfectly", "aligned", ".", "One", "hypothesis", "is", "that", "large", "image", "generative", "models", "face", "limitations", "when", "generating", "photorealistic", "human", "faces", "(", "Borji", ",", "2022", ")", ",", "and", "therefore", "some", "images", "generated", "with", "movie", "-", "related", "prompts", "appear", "to", "be", "closer", "to", "art", "and", "portrait", "regions", "in", "the", "embedding", "space", ".", "3.5", "Stable", "Diffusion", "Error", "Analysis", "We", "leverage", "DIFFUSION", "DBto", "discover", "Stable", "Dif-", "fusion", "generation", "failure", "cases", "and", "examine", "po-", "tential", "causes", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "To", "surface", "poor", "image", "generations", ",", "we", "compute", "CLIP", "embeddings", "for", "all", "prompts", "and", "images", "in", "DIFFUSION", "DB", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricValue", "I-MetricValue", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O" ]
[ "We", "then", "select", "prompt-", "image", "pairs", "with", "a", "large", "cosine", "distance", "(", "d", ")", "be-", "tween", "their", "embeddings", ".", "The", "cosine", "distances", "have", "a", "normal", "distribution", "(", "N", "(", "0.7123,0.04132", ")", ")", ".", "In", "this", "analysis", ",", "we", "focus", "on", "13,411", "β€œ", "bad", "”", "prompt-", "image", "pairs", "(", "1", ")", "with", "a", "distance", "that", "is", "larger", "than", "4", "standard", "deviations", "from", "the", "mean", "and", "(", "2", ")", "the", "image", "was", "not", "blurred", "by", "Stable", "Diffusion", "(", "Β§", "2.3", ")", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O" ]
[ "Impacts", "of", "hyperparameters", ".", "We", "conduct", "a", "lo-", "gistic", "regression", "test", "to", "analyze", "the", "relationship", "between", "Stable", "Diffusion", "hyperparameter", "values", "(", "e.g.", ",", "CFG", "scale", ",", "step", ",", "width", ",", "and", "height", ")", "and", "the", "likelihood", "of", "generating", "an", "image", "that", "is", "se-", "mantically", "different", "from", "its", "prompt", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "B-HyperparameterName", "I-HyperparameterName", "O", "B-HyperparameterName", "O", "B-HyperparameterName", "O", "O", "B-HyperparameterName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "The", "results", "reveal", "that", "all", "four", "hyperparameters", "are", "negatively", "correlated", "with", "the", "likelihood", "of", "generating", "a", "bad", "image", ".", "The", "correlation", "is", "statistically", "significant", "with", "a", "p", "-", "value", "of", "less", "than", "0.0001", "for", "all", "four", "vari-", "ables", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MetricName", "I-MetricName", "I-MetricName", "O", "O", "O", "B-MetricValue", "O", "O", "O", "O", "O", "O" ]
[ "Furthermore", ",", "we", "find", "the", "distribution", "of", "selected", "sampler", "options", "when", "generating", "bad", "im-", "ages", "is", "significantly", "different", "from", "the", "overall", "dis-", "tribution", "(", "X2=", "40873", ".11", ",", "p", "<", "0.0001", ")", ".", "CFG", "scale", "con-", "trols", "how", "much", "the", "generated", "image", "looks", "like", "the", "prompt", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterName", "I-HyperparameterName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "We", "find", "some", "users", "spec-", "ify", "negative", "CFG", "scales", "that", "make", "images", "look", "different", "from", "their", "prompts", "(", "large", "cosine", "distance", "d", ")", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterName", "I-HyperparameterName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "In", "the", "example", "shown", "on", "the", "right", ",", "a", "user", "gen-", "erates", "an", "image", "using", "a", "prompt", "about", "β€œ", "superman", "”", "with", "all", "default", "hyperparameters", "values", ",", "except", "for", "setting", "CFG", "scale", "to-1", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-HyperparameterName", "I-HyperparameterName", "B-HyperparameterValue", "O" ]
[ "This", "results", "in", "an", "image", "featuring", "a", "bowl", "of", "soup", "instead", "of", "β€œ", "superman", "”", ".", "A", "small", "step", "could", "also", "generate", "under-", "developed", "images", "that", "look", "different", "from", "the", "specified", "prompts", ".", "As", "demonstrated", "in", "the", "example", "on", "the", "right", ",", "a", "user", "generates", "an", "image", "about", "β€œ", "plague", "doctor", "”", "with", "all", "default", "hyperparameter", "values", ",", "except", "for", "setting", "step", "to2", ",", "which", "leads", "to", "a", "blurry", "image", ".", "Stable", "Diffusion", "struggles", "with", "generating", "images", "with", "a", "small", "size", "or", "large", "aspect", "ratios", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "The", "dissimilar", "image", "shown", "on", "the", "right", "is", "generated", "with", "default", "hyper-", "parameters", "except", "for", "a", "size", "of", "(", "64,512", ")", ".", "Impacts", "of", "prompts", ".", "Despite", "controlling", "all", "hyperparameters", "to", "be", "close", "to", "default", "values", ",", "we", "still", "find", "1.1k", "unique", "bad", "image", "-", "prompt", "pairs", ".", "Most", "of", "these", "instances", "have", "non", "-", "English", "prompts", ",", "very", "short", "prompts", ",", "or", "prompts", "consisting", "primarily", "emojis", "(", "see", "an", "ex-", "ample", "on", "the", "right", ")", ".", "The", "token", "lengths", "of", "these", "instances", "are", "significantly", "lower", "than", "the", "overall", "to-", "ken", "length", "(", "one", "-", "tailed", "t=βˆ’23.7203", ",", "p", "<", "0.0001", ")", ".", "The", "English", "prompt", "frequency", "among", "these", "in-", "stances", "is", "also", "significantly", "lower", "than", "the", "overall", "frequency", "(", "X2=", "1024", ".56", ",", "p", "<", "0.0001", ")", ".", "Inter-", "estingly", ",", "we", "also", "find", "that", "Stable", "Diffusion", "some-", "times", "generates", "unexpected", "images", "even", "when", "prompts", "are", "meaningful", "English", "sentences", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Future", "researchers", "can", "use", "our", "error", "analysis", "and", "failure", "cases", "to", "check", "potentially", "mislabeled", "training", "data", ".", "Implications", ".", "Our", "study", "reveals", "Stable", "Diffusion", "can", "make", "mistakes", "when", "generating", "images", "with", "certain", "hyperparameter", "values", "or", "prompt", "styles", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Negative", "CFG", "scales", ",", "small", "steps", ",", "or", "small", "sizes", "contributes", "to", "generating", "images", "dissimi-", "lar", "to", "prompts", "." ]
[ "O", "B-HyperparameterName", "I-HyperparameterName", "O", "O", "B-HyperparameterName", "O", "O", "O", "B-HyperparameterName", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Short", "and", "non", "-", "English", "prompts", "can", "also", "lead", "to", "errors", ".", "To", "improve", "the", "quality", "of", "fu-", "ture", "generative", "models", ",", "researchers", "can", "expand", "the", "training", "data", "to", "cover", "these", "edge", "cases", ".", "There", "are", "opportunities", "for", "researchers", "to", "design", "user", "inter-", "faces", "that", "can", "help", "users", "understand", "the", "impact", "of", "different", "hyperparameters", "and", "guide", "them", "in", "choos-", "ing", "values", "that", "fit", "their", "specific", "use", "cases", ".", "3.6", "Potentially", "Harmful", "Uses", "To", "identify", "potentially", "malicious", "uses", "of", "Stable", "Dif-", "fusion", ",", "we", "use", "named", "entity", "recognition", "to", "analyze", "prompts", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "We", "find", "that", "many", "prompts", "include", "names899", "of", "influential", "politicians", ",", "such", "as", "over", "65k", "images", "generated", "with", "a", "prompt", "including", "β€œ", "Donald", "Trump", "”", "and", "over", "48k", "images", "with", "β€œ", "Joe", "Biden", ".", "”", "Some", "prompts", "portray", "these", "politicians", "in", "negative", "lights", ",", "ranging", "from", "depicting", "them", "β€œ", "as", "Gollum", "with", "hair", "”", "to", "β€œ", "arrested", "in", "handcuffs", ".", "”", "Addition-", "ally", ",", "we", "find", "female", "celebrities", "are", "frequently", "used", "in", "prompts", ",", "with", "a", "high", "frequency", "after", "artists", "and", "influential", "politicians", ".", "Some", "of", "these", "prompts", "are", "presented", "in", "a", "sexual", "context", "that", "could", "be", "consid-", "ered", "nonconsensual", "pornography", ".", "Through", "keyword", "search", ",", "we", "discover", "prompts", "generating", "misinformation", "that", "could", "cause", "harm", ".", "For", "example", ",", "the", "prompt", "\"", "scientists", "putting", "microchips", "into", "a", "vaccine", "\"", "may", "harm", "pub-", "lic", "trust", "in", "medical", "institutions", "by", "potentially", "vali-", "dating", "conspiracy", "theories", ".", "Similarly", ",", "the", "prompt", "\"", "Russian", "soldiers", "in", "gas", "masks", "found", "the", "last", "surviving", "ukrainian", "after", "a", "nuclear", "war", "to", "liberate", "ukraine", "\"", "depicts", "false", "images", "of", "the", "Russo", "-", "Ukrainian", "War", "and", "could", "lead", "to", "new", "forms", "of", "propaganda", ".", "Our", "findings", "highlight", "the", "crucial", "need", "for", "further", "research", "on", "the", "broader", "impacts", "of", "large", "generative", "models", "and", "ways", "to", "regulate", "and", "mitigate", "their", "harms", ".", "4", "Enabling", "New", "Research", "Directions", "The", "unprecedented", "scale", "and", "diversity", "of", "DIFFU", "-", "SION", "DBbring", "new", "exciting", "research", "opportunities", "to", "help", "users", "generate", "images", "more", "effectively", "and", "efficiently", ",", "and", "enable", "researchers", "to", "improve", ",", "ex-", "plain", ",", "and", "safeguard", "generative", "models", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "I-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Prompt", "Autocomplete", "." ]
[ "B-TaskName", "I-TaskName", "O" ]
[ "With", "DIFFUSION", "DB", ",", "re-", "searchers", "can", "develop", "an", "autocomplete", "system", "to", "help", "users", "construct", "prompts", "." ]
[ "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "For", "example", ",", "one", "can", "use", "the", "prompt", "corpus", "to", "train", "an", "n", "-", "gram", "model", "to", "predict", "likely", "words", "following", "a", "prompt", "part", ".", "Alternatively", ",", "researchers", "can", "use", "semantic", "auto-", "complete", "(", "HyvΓΆnen", "and", "MΓ€kelΓ€", ",", "2006", ")", "by", "cate-", "gorizing", "prompt", "keywords", "into", "ontological", "cate-", "gories", "such", "as", "subject", ",", "style", ",", "quality", ",", "repetition", ",", "and", "magic", "terms", "(", "Oppenlaender", ",", "2022", ")", ".", "This", "allows", "the", "system", "to", "suggest", "related", "keywords", "from", "un-", "specified", "categories", ",", "for", "example", "suggesting", "style", "keyword", "β€œ", "depth", "of", "field", "”", "and", "a", "magic", "keyword", "β€œ", "award", "-", "winning", "”", "to", "improve", "the", "quality", "of", "gener-", "ated", "images", ".", "Additionally", ",", "researchers", "can", "also", "use", "DIFFUSION", "DBto", "study", "prompt", "auto", "-", "replace", "by", "distilling", "effective", "prompt", "patterns", "and", "creating", "a", "β€œ", "translation", "”", "model", "that", "replaces", "weaker", "prompt", "keywords", "with", "more", "effective", "ones", ".", "Generation", "through", "Search", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "AsDIFFUSION", "DB", "contains", "14", "million", "images", ",", "this", "dataset", "might", "have", "already", "included", "images", "with", "a", "user", "’s", "desired", "ef-", "fects", "." ]
[ "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Thus", ",", "a", "user", "can", "quickly", "search", "images", "in", "DIFFUSION", "DBinstead", "of", "running", "Stable", "Diffusion", ",", "which", "can", "be", "slow", "and", "costly", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Lexica", "(", "Shameem", ",", "2022", ")", ",", "an", "AI", "start", "-", "up", ",", "provides", "such", "a", "search", "en-", "gine", ",", "where", "users", "can", "search", "Stable", "Diffusion", "im-", "ages", "by", "natural", "language", "or", "images", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Researchers", "can", "also", "construct", "a", "structured", "index", "of", "images", "and", "prompts", ",", "such", "as", "building", "a", "semantivisual", "image", "hierarchy", "of", "images", "(", "Li", "et", "al", ".", ",", "2010", ")", "or", "a", "hierarchi-", "cal", "topic", "model", "of", "prompts", "(", "Griffiths", "et", "al", ".", ",", "2003", ")", ",", "to", "help", "users", "easily", "discover", "and", "explore", "images", "and", "prompts", "with", "similar", "styles", ".", "Improving", "Generative", "Models", ".", "With", "DIFFU", "-", "SION", "DB", ",", "a", "large", "and", "diverse", "collection", "of", "Sta-", "ble", "Diffusion", "usage", "logs", ",", "researchers", "not", "only", "can", "identify", "weak", "points", "and", "failure", "modes", "of", "Stable", "Diffusion", "but", "also", "gain", "insights", "into", "user", "pref-", "erences", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "I-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "For", "example", ",", "we", "demonstrate", "that", "re-", "searchers", "can", "use", "joint", "text", "-", "image", "embeddings", "be-", "tween", "prompts", "and", "images", "to", "detect", "generation", "mis-", "alignments", "(", "Β§", "3.5", ")", ".", "Additionally", ",", "DIFFUSION", "DB", "provides", "important", "metadata", "such", "as", "username", "hash", "andtimestamp", "for", "each", "generated", "image", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "By", "analyzing", "these", "metadata", "fields", ",", "researchers", "can", "trace", "the", "evolution", "chain", "of", "prompts", ",", "parameters", ",", "and", "images", ",", "which", "offers", "valuable", "insights", "into", "how", "users", "develop", "mental", "models", "of", "large", "generative", "models", "and", "their", "preferences", "of", "generated", "images", ".", "This", "understanding", "can", "inform", "future", "researchers", "to", "enhance", "generative", "models", "and", "design", "interfaces", "that", "facilitate", "better", "image", "-", "generation", "experiences", ".", "Explainable", "Generation", ".", "As", "generative", "models", "have", "been", "gaining", "immense", "popularity", ",", "there", "is", "a", "call", "for", "explainable", "creativity", "(", "Llano", "et", "al", ".", ",", "2022", ")", ".", "Many", "explanation", "techniques", "use", "input", "permuta-", "tion", "that", "computes", "feature", "attribution", "scores", "by", "running", "a", "model", "on", "slightly", "-", "modified", "input", "val-", "ues", "(", "Lundberg", "and", "Lee", ",", "2017", ")", ".", "DIFFUSION", "DB", "contains", "14", "million", "prompt", "-", "image", "pairs", "including", "similar", "prompts", "with", "minor", "differences", ",", "such", "as", "β€œ", "a", "happy", "dog", "”", "and", "β€œ", "a", "sad", "dog", "”", ",", "allowing", "re-", "searchers", "to", "investigate", "how", "individual", "keywords", "affect", "the", "generation", "process", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Deepfake", "Detection", "." ]
[ "B-TaskName", "I-TaskName", "O" ]
[ "Breakthroughs", "in", "gener-", "ative", "models", "raise", "concerns", "about", "deepfakes", "β€”", "fake", "images", "of", "real", "individuals", "for", "unethical", "pur-", "poses", "(", "Wiggers", ",", "2022", ")", ".", "DIFFUSION", "DBis", "valu-900", "able", "for", "detecting", "deepfakes", ",", "as", "it", "contains", "a", "large-", "scale", "collection", "of", "model", "-", "generated", "images", "and", "their", "metadata", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "B-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Researchers", "can", "use", "this", "collection", "to", "train", "ML", "models", "to", "identify", "synthetic", "artifacts", "and", "train", "classifiers", "that", "classify", "synthetic", "images", "from", "real", "images", "(", "Mirsky", "and", "Lee", ",", "2022", ")", ".", "5", "Related", "Work", "Text", "-", "to", "-", "text", "Prompting", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "O" ]
[ "Researchers", "have", "been", "studying", "prompt", "engineering", "for", "text", "-", "to", "-", "text", "gener-", "ation", "(", "e.g.", ",", "Liu", "et", "al", ".", ",", "2022", ";", "Lu", "et", "al", ".", ",", "2022", ";", "Rubin", "et", "al", ".", ",", "2022", ")", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "To", "facilitate", "this", "line", "of", "research", ",", "researchers", "develop", "PromptSource", "(", "Bach", "et", "al", ".", ",", "2022", ")", ",", "a", "dataset", "of", "2k", "text", "prompts", "along", "with", "a", "framework", "to", "create", "and", "share", "prompts", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "In", "contrast", ",", "our", "work", "focuses", "on", "text", "-", "to", "-", "image", "prompting", ",", "and", "DIFFUSION", "DBhas", "an", "unprecedented", "scale", "of", "14", "million", "real", "prompt", "-", "image", "pairs", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "Text", "-", "to", "-", "image", "Prompting", "." ]
[ "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "O" ]
[ "There", "is", "a", "growing", "interest", "in", "text", "-", "to", "-", "image", "prompt", "engineering", "re-", "search", "from", "NLP", ",", "Computer", "Vision", ",", "and", "HCI", "communities", "(", "e.g.", ",", "Qiao", "et", "al", ".", ",", "2022", ";", "Pavlichenko", "and", "Ustalov", ",", "2022", ")", "." ]
[ "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "For", "example", ",", "Oppenlaen-", "der", "(", "2022", ")", "identifies", "six", "types", "of", "prompt", "modi-", "fiers", "through", "an", "ethnographic", "study", ",", "and", "Liu", "and", "Chilton", "(", "2022", ")", "proposes", "design", "guidelines", "for", "text-", "to", "-", "image", "prompt", "engineering", "by", "experimenting", "with", "1,296", "prompts", "." ]
[ "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "I-TaskName", "O", "O", "O", "O", "O", "O" ]
[ "Closest", "in", "spirit", "to", "DIFFU", "-", "SION", "DBis", "Lexica", "(", "Shameem", ",", "2022", ")", "which", "allows", "users", "to", "search", "over", "5", "million", "Stable", "Diffusion", "im-", "ages", "with", "their", "prompts", ",", "but", "it", "does", "not", "release", "its", "internal", "database", "." ]
[ "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "I-DatasetName", "I-DatasetName", "B-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "In", "comparison", ",", "DIFFUSION", "DB", "is", "open", "-", "source", "and", "publicly", "available", "to", "everyone", "." ]
[ "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O" ]
[ "6", "Conclusion", "We", "present", "DIFFUSION", "DB", ",", "the", "first", "large", "-", "scale", "text", "-", "to", "-", "image", "prompt", "dataset", ",", "containing", "14", "million", "images", "with", "their", "prompts", "and", "hyperparameters", "col-", "lected", "from", "the", "Stable", "Diffusion", "discord", "server", "." ]
[ "O", "O", "O", "O", "B-DatasetName", "I-DatasetName", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MethodName", "I-MethodName", "O", "O", "O" ]