Molweni: A Challenge Multiparty Dialogues-based Machine Reading Comprehension Dataset with Discourse Structure

Research into the area of multiparty dialog has grown considerably over\nrecent years. We present the Molweni dataset, a machine reading comprehension\n(MRC) dataset with discourse structure built over multiparty dialog. Molweni's\nsource samples from the Ubuntu Chat Corpus, including 10,000 dialogs comprising\n88,303 utterances. We annotate 30,066 questions on this corpus, including both\nanswerable and unanswerable questions. Molweni also uniquely contributes\ndiscourse dependency annotations in a modified Segmented Discourse\nRepresentation Theory (SDRT; Asher et al., 2016) style for all of its\nmultiparty dialogs, contributing large-scale (78,245 annotated discourse\nrelations) data to bear on the task of multiparty dialog discourse parsing. Our\nexperiments show that Molweni is a challenging dataset for current MRC models:\nBERT-wwm, a current, strong SQuAD 2.0 performer, achieves only 67.7% F1 on\nMolweni's questions, a 20+% significant drop as compared against its SQuAD 2.0\nperformance.\n