Datasets:

havens2

/

naacl2022

like 0

[ "Here", "we", "see", "that", "most", "systems", "struggle", "with", "the", "minority", "class", "(“F”),", "but", "our", "combined", "approach", "appears", "to", "handle", "this", "well." ]

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

[ "Another", "thing", "of", "note", "is", "CoAID,", "the", "only", "dataset", "where", "the", "class", "distribution", "is", "heavily", "imbalanced." ]

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

[ "use", "the", "user", "network", "(RNN+CNN", "and", "GCAN),", "although", "the", "strong", "performance", "of", "DUCK", "indicates", "that", "combining", "both", "types", "of", "information", "works", "best,", "suggesting", "that", "they", "complement", "each", "other." ]

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

[ "“CT”,", "“CC”", "and", "“UT”", "denote", "comment", "tree,", "comment", "chain", "and", "user", "tree", "respectively,", "and", "“R”", "and", "“NR”", "in", "WEIBO", "denote", "rumour", "and", "non-rumour." ]

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

[ "10https://github.com/majingCUHK/Rumor_R" ]

[ 0 ]

[ "followers,", "and", "so", "it", "uses", "GATprf." ]

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

[ "9With", "the", "exception", "of", "WEIBO", "where", "we", "can’t", "crawl", "users’" ]

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

[ "We", "also", "observe", "that", "models", "that", "use", "the", "comment", "texts", "(stance-BERT", "and", "Bi-GCN)", "tend", "to", "do", "better", "than", "those", "that", "only" ]

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

[ "In", "terms", "of", "datasets,", "WEIBO", "appears", "to", "be", "the", "“easier”", "dataset,", "where", "most", "systems", "produce", "a", "macro-F1", "over", "90%." ]

[ 0, 0, 0, 0, 13, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 7, 0, 9 ]

[ "DUCK", "(our", "model)", "performs", "very", "strongly,", "outperforming", "all", "benchmark", "systems", "consistently", "over", "all", "datasets,", "creating", "a", "new", "state-of-the-art", "for", "rumour", "detection." ]

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

[ "We", "present", "the", "results", "in", "Table", "5." ]

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

[ "Note", "that", "we", "only", "have", "English", "results", "(Twitter15,", "Twitter16", "and", "CoAID)", "for", "stanceBERT", "as", "it", "uses", "stance", "annotations", "from", "SemEval2016", "(Mohammad", "et", "al.,", "2016),", "and", "GCAN", "and", "RNN+CNN", "do", "not", "have", "results", "for", "CoAID", "as", "it", "does", "not", "contain", "retweets." ]

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

[ "All", "benchmark", "results", "are", "produced", "using", "the", "author-provided", "code,", "with", "the", "exception", "of", "RNN+CNN", "and", "stance-BERT", "where", "we", "implement", "ourselves." ]

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

[ "For", "a", "summary", "of", "the", "different", "features", "used", "by", "these", "benchmark", "systems", "and", "our", "model,", "see", "Table", "1." ]

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

[ "We", "next", "compare", "the", "rumour", "detection", "performance", "of", "DUCK", "that", "combines", "comment", "tree,", "comment", "chain", "and", "user", "tree", "models", "(Figure", "1)", "to", "the", "following", "state-of-the-art", "methods:", "(1)", "RvNN", "(Ma", "et", "al.,", "2018)10:", "uses", "a", "GRU", "to", "process", "text", "content", "and", "recursive", "networks", "to", "model", "the", "comment", "network;", "(2)", "RNN+CNN", "(Liu", "and", "Wu,", "2018):", "uses", "CNN", "and", "RNN", "to", "model", "the", "retweet", "user", "network", "where", "user", "representations", "are", "initialised", "with", "user", "profile", "features;", "(3)", "stance-BERT", "(Tian", "et", "al.,", "2020):", "fine-tunes", "a", "BERT", "pretrained", "with", "stance", "annotations", "for", "rumour", "detection", "and", "comments", "are", "modelled", "as", "a", "chain", "(similar", "to", "our", "one-tier", "transformer", "model);", "(4)", "Bi-GCN", "(Bian", "et", "al.,", "2020)11:", "uses", "a", "bidirectional", "graph", "convolutional", "network", "to", "model", "the", "comment", "network", "in", "a", "top-down", "(i.e.", "nodes", "are", "combined", "starting", "from", "the", "leaf", "comments)", "and", "bottom-up", "manner", "(i.e.", "nodes", "are", "combined", "starting", "from", "the", "root);", "and", "(5)", "GCAN", "(Lu", "and", "Li,", "2020)12:", "uses", "graph", "networks", "to", "model", "the", "retweet", "user", "network", "and", "a", "CNN", "to", "model", "the", "source", "post", "with", "co-attention", "between", "the", "two", "networks." ]

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

[ "5.2.4", "Overall", "Rumour", "Detection" ]

[ 0, 0, 11, 12 ]

[ "reply", "and", "retweet", "user", "network.9" ]

[ 0, 0, 0, 0, 0 ]

[ "8For", "one-tier", "and", "two-tier", "transformers,", "if", "the", "number", "of", "comments", "is", "set", "to", "10,", "that", "means", "we", "will", "concatenate", "10", "comments", "(with", "the", "source", "post)", "into", "a", "long", "string,", "and", "any", "text", "that", "exceeds", "BERT’s", "maximum", "sequence", "length", "will", "be", "truncated", "(and", "so", "for", "some", "stories", "the", "models", "may", "use", "less", "than", "10", "comments,", "if", "earlier", "comments", "are", "very", "long)." ]

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

[ "Our", "results", "highlight", "the", "importance", "of", "incorporating", "social", "relations,", "and", "DUCK", "therefore", "uses", "GATprf+rel", "for", "modelling", "the" ]

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

[ "Unsurprisingly,", "random", "initialisation", "performs", "the", "worst,", "and", "we", "see", "a", "substantial", "improvement", "when", "user", "profile", "information", "is", "incorporated,", "and", "again", "an", "improvement", "when", "we", "incorporate", "user", "social", "relations", "(6%", "and", "5%", "increase", "in", "macro-F1", "for", "Twitter16", "and", "CoAID)." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 9, 0, 9, 0, 7, 8, 0, 13, 0, 13 ]

[ "Results", "are", "shown", "in", "Table", "4." ]

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

[ "Recall", "that", "we", "use", "GAT", "to", "represent", "the", "reply", "and", "repost", "user", "network,", "and", "we", "investigate", "different", "node", "encodings", "to", "initialise", "GAT:", "(1)", "random", "initialisation", "(GATrnd);", "(2)", "user", "profile", "metadata", "(GATprf);", "and", "(3)", "user", "profile", "metadata", "and", "social", "relations", "(GATprf+rel)." ]

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

[ "5.2.3", "User", "Tree" ]

[ 0, 0, 0 ]

[ "With", "these", "results,", "we", "will", "use", "the", "two-tier", "transformer", "to", "model", "the", "comment", "chain", "in", "DUCK." ]

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

[ "Interestingly,", "even", "though", "longformer", "is", "able", "to", "include", "most", "of", "the", "comments,", "it", "performs", "worse", "than", "both", "one-tier", "and", "two-tier", "transformer,", "suggesting", "that", "the", "sparser", "attention", "pattern", "that", "longformer", "introduces", "has", "a", "negative", "impact." ]

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

[ "Two-tier", "transformer", "is", "able", "to", "process", "more", "comments,", "and", "produces", "the", "best", "performance." ]

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

[ "However,", "due", "to", "one-tier", "transformer’s", "sequence", "length", "limit,", "it", "can", "take", "no", "more", "than", "60", "comments", "on", "average." ]

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

[ "Both", "one-tier", "and", "two-tier", "transformers", "see", "a", "performance", "gain", "when", "the", "number", "of", "comments", "increases", "and", "a", "drop", "when", "there", "are", "too", "many", "comments", "(noting", "that", "the", "trend", "is", "flatter", "in", "CoAID)." ]

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

[ "7The", "“unpaired”", "approach", "uses", "“all”", "as", "the", "aggregation" ]

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

[ "5.2.2", "Comment", "Chain", "Recall", "that", "we", "explore", "using", "transformer", "models", "–", "one-tier", "transformer,", "longformer", "and", "two-tier", "trans" ]

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

[ "Given", "these", "results,", "our", "full", "model", "(DUCK)", "will", "be", "using", "“all”", "as", "the", "aggregation", "method", "for", "computing", "the", "comment", "graph", "representation." ]

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

[ "The", "answer", "is", "evidently", "yes,", "as", "we", "see", "a", "substantial", "drop", "in", "performance", "when", "we", "process", "the", "posts", "independently:", "“unpaired”", "produces", "a", "macro-F1", "of", "only", "0.83", "in", "both", "Twitter16", "and", "CoAID." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 7, 0, 0, 9, 0, 0, 13, 0, 13 ]

[ "Does", "processing", "CoAID", "drops", "to", "0.80", "with", "the", "parent-child", "posts", "together", "with", "BERT", "help?" ]

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

[ "We", "can", "see", "that", "the", "root", "and", "its", "immediate", "neighbours", "contain", "most", "of", "the", "information,", "and", "not", "including", "the", "root", "node", "impacts", "the", "performance", "severely", "(both", "Twitter16", "and", "root)." ]

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

[ "Comparing", "the", "aggregation", "methods,", "“all”", "performs", "the", "best,", "followed", "by", "“N”", "and", "“root”", "(0.88", "vs.", "0.87", "vs.", "0.86", "in", "Twitter16;", "0.87", "vs.", "0.86", "vs.", "0.85", "in", "CoAID", "in", "terms", "of", "Macro-F1)." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 9, 0, 9, 0, 9, 0, 13, 9, 0, 9, 0, 9, 0, 13, 0, 0, 0, 7 ]

[ "We", "report", "the", "performance", "of", "this", "alternative", "model", "(“unpaired”)7", "and", "the", "different", "aggregation", "methods", "(“root”,", "“", "root”,", "“N”", "and", "“all”;", "equations", "3,", "4,", "5", "and", "6", "respectively)", "in", "Table", "3." ]

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

[ "That", "is,", "Equation", "1", "is", "now", "modified", "to:" ]

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

[ "5.2.1", "Comment", "Tree", "To", "understand", "the", "impact", "of", "using", "BERT", "for", "processing", "a", "pair", "of", "parent-child", "posts,", "we", "present", "an", "alternative", "method", "(“unpaired”)", "where", "we", "use", "BERT", "to", "process", "each", "post", "independently", "before", "feeding", "their", "[CLS]", "representation", "to", "the", "GAT." ]

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

[ "All", "results", "are", "an", "average", "over", "5", "runs", "(5-fold", "cross-validation", "for", "Twitter15/16", "and", "5", "independent", "runs", "with", "different", "random", "seeds", "for", "WEIBO", "and", "CoAID", "following", "Ma", "et", "al.", "(2016,", "2017);", "Cui", "and", "Lee", "(2020))." ]

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

[ "In", "terms", "of", "evaluation", "metrics,", "we", "present", "F1", "scores", "for", "each", "class", "and", "macro-averaged", "F1", "scores", "as", "the", "aggregate", "performance." ]

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

[ "For", "the", "first", "three", "questions,", "we", "present", "development", "performance", "using", "Twitter16", "and", "CoAID", "as", "the", "representative", "datasets", "(as", "the", "trends", "are", "largely", "the", "same", "for", "the", "other", "datasets),", "while", "for", "the", "final", "question", "we", "report", "the", "test", "performance", "for", "all", "datasets." ]

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

[ "[Overall", "performance]:", "Do", "the", "three", "different", "components", "complement", "each", "other", "and", "how", "does", "a", "combined", "approach", "compared", "to", "existing", "rumour", "detection", "systems?" ]

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

[ "•", "Q3", "[User", "tree]:" ]

[ 0, 0, 0, 0 ]

[ "•", "Q2", "[Comment", "chain]:", "Does", "incorporating", "more", "comments", "help", "rumour", "detection", "when", "modelling", "them", "as", "a", "stream", "of", "posts?" ]

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

[ "tree]:", "Does", "incorporating", "BERT", "to", "analyse", "the", "relation", "between", "parent", "and", "child", "posts", "help", "modelling", "the", "comment", "network,", "and", "what", "is", "the", "best", "way", "to", "aggregate", "comment-pair", "encodings", "to", "represent", "the", "comment", "graph?" ]

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

[ "Specifically,", "we", "are", "interested", "in", "the", "following", "questions:" ]

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

[ "For", "the", "first", "set", "of", "results", "where", "we", "evaluate", "each", "module", "independently,", "we", "feed", "their", "representations", "(i.e.", "zct,", "zcc", "and", "zut)", "to", "an", "MLP", "layer", "to", "do", "classification." ]

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

[ "We", "first", "present", "results", "for", "each", "of", "the", "modules", "(comment", "tree,", "comment", "chain", "and", "user", "tree)", "separately", "to", "understand", "the", "best", "approach", "for", "modelling", "them,", "and", "then", "present", "the", "final", "results", "where", "we", "compare", "our", "full", "model", "DUCK", "to", "a", "number", "of", "benchmark", "systems." ]

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

[ "5.2", "Results" ]

[ 0, 0 ]

[ "6For", "Twitter15/16", "during", "tuning", "we", "use", "only", "one", "of", "the", "folds", "and", "reserve", "1/4", "of", "the", "training", "data", "as", "development", "set", "and", "train", "the", "model", "using", "the", "rest", "(3/4)", "of", "the", "training", "data." ]

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

[ "s/twitter-api/v1" ]

[ 0 ]

[ "5https://developer.twitter.com/en/doc" ]

[ 0 ]

[ "In", "this", "section,", "we", "first", "introduce", "the", "datasets", "for", "our", "experiments", "and", "then", "present", "the", "performance", "of", "DUCK", "by", "comparing", "against", "a", "number", "state-ofthe-art", "models." ]

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

[ "Additional", "implementation", "details", "for", "all", "models" ]

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

[ "We", "use", "the", "development", "set", "of", "each", "dataset", "for", "tuning", "hyper-parameters.6" ]

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

[ "For", "CoAID", "and", "WEIBO,", "we", "reserve", "20%", "data", "as", "test", "and", "split", "the", "rest", "in", "a", "ratio", "of", "4:1", "for", "training", "and", "development", "partitions", "and", "report", "the", "average", "test", "performance", "over", "5", "runs", "(initialised", "with", "different", "random", "seeds)." ]

[ 0, 13, 0, 13, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 5, 0, 3, 4, 4, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "In", "terms", "of", "data", "partitioning,", "for", "Twitter15", "and", "Twitter16", "we", "follow", "previous", "studies", "(Ma", "et", "al.,", "2015,", "2016)", "and", "report", "the", "average", "performance", "based", "on", "5-fold", "cross-validation." ]

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

[ "For", "Twitter-based", "datasets", "(Twitter15/16", "and", "CoAID),", "we", "crawl", "the", "tweets", "and", "additional", "user", "information", "(e.g.", "user", "profile", "metadata", "and", "followers)", "via", "the", "official", "Twitter", "API.5", "For", "WEIBO,", "the", "platform", "does", "not", "provide", "a", "means", "to", "crawl", "social", "relations", "and", "so", "the", "user", "tree", "uses", "GATprf." ]

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

[ "Table", "2", "presents", "some", "statistics", "of", "these", "datasets." ]

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

[ "WEIBO", "(Ma", "et", "al.,", "2016)", "contains", "a", "collection", "of", "stories", "from", "Sina", "Weibo,", "a", "Chinese", "social", "media", "platform,", "and", "is", "annotated", "with", "two", "classes", "(rumour", "and", "non-rumour)." ]

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

[ "CoAID", "(Cui", "and", "Lee,", "2020)", "collects", "of", "a", "set", "of", "COVID-19", "news", "articles", "shared", "on", "Twitter,", "and", "they", "are", "annotated", "with", "two", "classes", "(true", "or", "fake)." ]

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

[ "Twitter15", "and", "Twitter16", "are", "Twitter", "datasets", "with", "four", "rumour", "classes:", "true", "rumour,", "false", "rumour,", "non-rumour", "and", "unverified", "rumour." ]

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

[ "We", "evaluate", "our", "method", "on", "four", "widely", "used", "rumour", "datasets:", "Twitter15", "(Ma", "et", "al.,", "2017);", "Twitter16", "(Ma", "et", "al.,", "2017);", "CoAID", "(Cui", "and", "Lee,", "2020);", "and", "WEIBO", "(Ma", "et", "al.,", "2016)." ]

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

[ "5", "Experiments", "and", "Results" ]

[ 0, 0, 0, 0 ]

[ "where", "n", "denotes", "the", "number", "of", "training", "instances." ]

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

[ "Given", "an", "optimal", "approach", "for", "each", "module", "(Section", "5),", "DUCK", "combines", "the", "output", "from", "all", "modules", "to", "classify", "a", "source", "post", "and", "is", "trained", "using", "standard", "crossentropy", "loss." ]

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

[ "In", "each", "module", "(comment", "tree,", "comment", "chain", "and", "user", "tree),", "we", "explore", "a", "number", "of", "approaches", "to", "model", "its", "structure", "(e.g.", "there", "are", "several", "ways", "to", "aggregate", "the", "node", "encodings", "to", "produce", "zct", "for", "the", "comment", "tree", "and", "3", "different", "methods", "to", "produce", "zcc", "for", "the", "comment", "chain)." ]

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

[ "The", "output", "of", "our", "decoder", "is", "a", "reconstructed", "adjacency", "matrix", "ˆA.", "Formally:" ]

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

[ "The", "decoder", "is", "defined", "by", "an", "inner", "product", "between", "latent", "variable", "Z." ]

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

[ "It", "takes", "an", "adjacency", "matrix", "A", "and", "a", "feature", "matrix", "X", "as", "inputs", "and", "generates", "the", "latent", "variable", "Z", "as", "output." ]

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

[ "We", "use", "a", "two-layer", "GCN", "as", "the", "encoder." ]

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

[ "Our", "goal", "is", "to", "learn", "a", "transformad,", "which", "converts", "users", "to", "a", "tion", "matrix", "Z", "latent", "space", "with", "the", "dimensionality", "of", "d." ]

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

[ "Given", "the", "social", "graph", "Gs", "constructed", "based", "on", "the", "training", "data,", "we", "can", "derive", "an", "adjacency", "man,", "where", "n", "is", "the", "number", "of", "users." ]

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

[ "4944" ]

[ 0 ]

[ "4For", "the", "unseen", "or", "isolated", "users,", "we", "initialise", "them", "based", "on", "their", "user", "features", "(used", "in", "GATprf),", "and", "project", "them", "via", "a", "learned", "matrix", "into", "the", "same", "space", "as", "the", "GAE-initialised", "user", "nodes." ]

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

[ "3The", "last", "feature", "is", "technically", "not", "user", "profile", "information,", "but", "it", "is", "a", "form", "of", "user", "characteristic", "as", "it", "captures", "how", "quickly", "they", "engage", "with", "a", "post." ]

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

[ "The", "idea", "for", "using", "GAE-learned", "encodings", "to", "initialise", "user", "nodes", "is", "that", "they", "are", "more", "informative,", "since", "they", "capture", "information", "about", "a", "user", "and", "their", "peers." ]

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

[ "Note", "that", "the", "network", "structure", "of", "the", "GAT", "and", "GAE", "is", "intrinsically", "different", "—", "the", "former", "captures", "the", "users", "that", "engage", "with", "a", "source", "post", "while", "the", "latter", "the", "network", "of", "users", "who", "follow", "one", "another." ]

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

[ "Intuitively,", "GAE", "is", "an", "unsupervised", "graph", "learning", "algorithm", "that", "takes", "in", "an", "adjacency", "matrix", "as", "input,", "and", "learns", "node", "representations", "that", "can", "reconstruct", "the", "adjacency", "matrix", "in", "the", "output." ]

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

[ "GATprf+rel:", "This", "method", "initialises", "the", "user", "nodes", "with", "representations", "learned", "by", "a", "variational", "graph", "autoencoder", "(GAE;", "Kipf", "and", "Welling", "(2016))", "based", "on", "the", "user", "features", "(defined", "above)", "and", "their", "social", "relations", "(based", "on", "“follow”", "relations).4" ]

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

[ "GATprf:", "Following", "Liu", "and", "Wu", "(2018),", "this", "method", "initialises", "the", "user", "nodes", "based", "on", "features", "derived", "from", "their", "user", "profiles:", "username,", "user", "screen", "name,", "user", "description,", "user", "account", "age,", "number", "of", "followers", "and", "following", "users,", "number", "of", "posts", "and", "favourite", "posts,", "whether", "the", "profile", "is", "protected,", "whether", "the", "account", "is", "GPS-enabled,", "and", "the", "time", "difference", "with", "the", "source", "post.3", "Thus,", "the", "static", "user", "node", "h0", "Rk", "from", "user", "profiles" ]

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

[ "tialise", "the", "user", "nodes", "with", "random", "vectors." ]

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

[ "GATrnd:", "This", "is", "the", "base", "method", "where", "we", "ini" ]

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

[ "i" ]

[ 0 ]

[ "):" ]

[ 0 ]

[ "is", "in", "how", "they", "initialise", "the", "user", "nodes", "(h(0)" ]

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

[ "The", "main", "difference", "between", "the", "three", "methods" ]

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

[ "2018)", "to", "model", "the", "network", "(following", "Equation", "2),", "and", "we", "aggregate", "the", "node", "encodings", "by", "meanpooling", "over", "all", "nodes", "to", "produce", "the", "graph", "representation:" ]

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

[ "model_doc/longformer.html" ]

[ 0 ]

[ "2https://huggingface.co/transformers/" ]

[ 0 ]

[ "All", "methods", "use", "a", "GAT", "(Veliˇckovi´c", "et", "al.," ]

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

[ "We", "explore", "three", "methods", "to", "model", "the", "user", "network." ]

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

[ "Note", "that", "unlike", "previous", "studies,", "our", "user", "network", "captures", "all", "users", "who", "reply", "to", "or", "repost", "the", "source", "post", "(previous", "studies", "use", "only", "the", "reposts,", "see", "Table", "1)." ]

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

[ "Previous", "studies", "found", "that", "the", "user", "characteristics", "are", "different", "for", "those", "that", "engage", "with", "rumours", "vs.", "those", "who", "don’t", "(Vosoughi", "et", "al.,", "2018;", "Shu", "et", "al.,", "2018),", "motivating", "us", "to", "model", "the", "user", "network." ]

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

[ "Moving", "away", "from", "the", "post", "content,", "here", "we", "model", "the", "network", "of", "users", "that", "interact", "with", "a", "source", "post", "(“user", "tree”", "in", "Figure", "1)." ]

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

[ "4.3", "User", "Tree" ]

[ 0, 0, 0 ]

[ "The", "secondtier", "transformer", "has", "a", "similar", "architecture", "to", "BERT,", "but", "has", "only", "2", "layers", "and", "its", "parameters", "are", "initialised", "randomly." ]

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

[ "where", "BERT", "and", "transformer", "denote", "the", "first-", "and", "second-tier", "transformers", "respectively." ]

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

[ "We", "use", "a", "pretrained", "longformer2,", "and", "follow", "the", "same", "approach", "as", "before", "for", "modelling", "the", "comment", "chain:" ]

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

[ "Longformer", "has", "a", "similar", "architecture", "as", "the", "onetier", "transformer,", "but", "uses", "a", "sparser", "attention", "pattern", "to", "process", "longer", "sequences", "more", "efficiently." ]

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