Early Rumor Detection Using Neural Hawkes Process with a New Benchmark Dataset

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4105 - 4117 July 10-15, 2022 ©2022 Association for Computational Linguistics Early Rumor Detection Using Neural Hawkes Process with a New Benchmark Dataset Fengzhu Zeng Wei Gao Singapore Management University Singapore Management University 81 Victoria St, Singapore 188065 81 Victoria St, Singapore 188065 fzzeng.2020@phdcs.smu.edu.sg weigao@smu.edu.sg Abstract Little attention has been paid on EArly Rumor Detection (EARD), and EARD performance was evaluated inappropriately on a few datasets where the actual early-stage information is largely missing. To reverse such situation, we construct BEARD, a new Benchmark dataset for EARD, based on claims from fact-checking websites by trying to gather as many early relevant posts as possible. We also pro- pose HEARD, a novel model based on neural Hawkes process for EARD, which can guide a generic rumor detection model to make timely, accurate and stable predictions. Experiments show that HEARD achieves effective EARD performance on two commonly used general ru- mor detection datasets and our BEARD dataset. 1 Introduction The proliferation of online rumors has aroused widespread concerns. Many studies have been con- ducted for automatic rumor detection in social me- dia and achieved high detection accuracy (Ma et al., 2016, 2017; Yu et al., 2017; Ruchansky et al., 2017; Ma et al., 2018; Guo et al., 2018; Bian et al., 2020). However, these generic detection methods lack in- depth modeling of temporality, which can cause tremendous delay in detection given the instanta- neous and sporadic nature of rumor propagation. While a few EArly Rumor Detection (EARD) mod- els have been proposed (Liu and Wu, 2018; Zhou et al., 2019; Song et al., 2019; Xia et al., 2020), they have been designed with oversimplifcation and evaluated inappropriately using the datasets con- structed for generic rumor detection. Widely used rumor detection datasets, such as TWITTER (Ma et al., 2016) and PHEME (Zubiaga et al., 2016), are generally limited in covering relevant posts in the early stage as there was no mechanism ensur- ing to gather information that is further away from the offcial debunking time of a rumor. For this reason, the generalizability of EARD cannot be Time Post 2015-07-07 Translucent butterfy - beautiful! T.ORG. 2015-10-01 #Snopes Translucent Butterfy URL 2013-02-08 Ever see a translucent butterfy? 2013-06-24 fake...like bubbles T.REC. 2014-09-03 ...fake, a Worth1000 Photoshop contest entry 2014-10-06 Multiple repeat fakes 01-08 00:00 . . . courtesy of Banksy. P.ORG. 01-08 14:39 ...it is not by Banksy its by @LucilleClerc 01-07 20:09 Banksy for Charlie... 01-07 23:48 That’s not Banksy though, just someone fan page P.REC. 01-08 00:14 . . . I don’t think that Banksy Insta is offcial is it? 01-08 08:33 The person (not B.) shared it from another source. 01-08 09:55 Not Banksy btw. It’s @LucilleClerc Table 1: A rumor in TWITTER dataset claiming "A but- terfy with translucent wings" and another from PHEME claiming "A street artist Banksy posted an illustration on the Instagram as tribute to Charlie Hebdo". …

Content truncated to 500 words