Weakly supervised discourse segmentation for multiparty oral conversations

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1381–1392 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 1381 Weakly supervised discourse segmentation for multiparty oral conversations Lila Gravellier,1 Julie Hunter,2 Philippe Muller,1,3 Thomas Pellegrini,1,3 Isabelle Ferrané1 1IRIT, University of Toulouse, 2LINAGORA Labs, 3Artificial and Natural Intelligence Toulouse Institute (ANITI) firstname.lastname@irit.fr, jhunter@linagora.com Abstract Discourse segmentation, the first step of dis- course analysis, has been shown to improve results for text summarization, translation and other NLP tasks. While segmentation mod- els for written text tend to perform well, they are not directly applicable to spontaneous, oral conversation, which has linguistic features for- eign to written text. Segmentation is less stud- ied for this type of language, where annotated data is scarce, and existing corpora more het- erogeneous. We develop a weak supervision approach to adapt, using minimal annotation, a state of the art discourse segmenter trained on written text to French conversation transcripts. Supervision is given by a latent model boot- strapped by manually defined heuristic rules that use linguistic and acoustic information. The resulting model improves the original seg- menter, especially in contexts where informa- tion on speaker turns is lacking or noisy, gain- ing up to 13% in F-score. Evaluation is per- formed on data like those used to define our heuristic rules, but also on transcripts from two other corpora. 1 Introduction Discourse analysis, focusing on pragmatic aspects of text interpretation, especially aspects ranging beyond the level of the sentence, is a long standing domain of Natural Language Processing (NLP), of growing importance for many tasks in NLP such as Machine Translation (MT) (Chen et al., 2020) and summarization (Louis et al., 2010; Xu et al., 2020). The first level of discourse analysis consists in segmenting a discourse into basic units, which gen- erally correspond to roughly clause-level units of text whose contents provide the arguments to dis- course relations such as Explanation, Elaboration, and Contrast (Mann and Thompson, 1987; Prasad et al., 2008; Asher and Lascarides, 2003). Tak- ing the resulting discourse segments as input for tasks such as summarization/sentence compression can improve performance over sentence-based ap- proaches (Li et al., 2020; Sporleder and Lapata, 2005; Xu et al., 2020). Segmentation is also cru- cial for downstream tasks that exploit discourse relations (Chen et al., 2020; Louis et al., 2010; Xu et al., 2020). Because automatic segmentation on text-based corpora tends to yield good results (above 90% F- score on segment boundary recognition, depending on the language, cf. Zeldes et al., 2019), the seg- mentation task is often neglected in favor of the prediction of discourse relations. State of the art discourse segmenters, however, generally take sen- tence boundaries as given, and often benefit from other forms of punctuation, as well as the well- formed sentences found in text-based corpora. Such systems provide a less stable foundation for discourse parsing of spoken conversations, which is necessary to improve, e.g., real time recommen- dations from voice assistants and meeting summa- rization, and to develop …

Content truncated to 500 words