A simple but effective model for attachment in discourse parsing with multi-task learning for relation labeling

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 3412–3417 May 2-6, 2023 ©2023 Association for Computational Linguistics A simple but effective model for attachment in discourse parsing with multi-task learning for relation labeling Zineb Bennis IRIT zinebennisa@gmail.com Julie Hunter Linagora jhunter@linagora.com Nicholas Asher IRIT Nicholas.Asher@irit.fr Abstract We present a discourse parsing model for conver- sation trained on the STAC corpus (Asher et al., 2016). We fine-tune a BERT-based model to en- code pairs of discourse units and use a simple lin- ear layer to predict discourse attachments. We then exploit a multi-task setting to predict relation la- bels, which effectively aids in the difficult task of relation type prediction; our F1-score equals or sur- passes the state of the art in the approaches we have reimplemented using code from the authors with no loss in performance for attachment, confirming the intuitive interdependence of these two tasks. Our method also improves over other discourse parsing models in the literature in permitting attachments in which one node has multiple parents, an important feature of multiparty conversation. 1 Introduction Discourse parsing, the task of predicting graphs that represent semantic relations (arcs) between elementary discourse units or EDUs (nodes), is a hard problem in NLP due to the complexity of discourse graphs and the frequent lack of surface cues provided by EDUs, which forces parsers to rely on deep, semantic information. Multiparty, spontaneous conversation is especially tricky as the structure often meanders and a single participant can respond to multiple discourse moves at once, yielding non-tree-like structures that challenge ex- istent parsing techniques (Afantenos et al., 2015). While it is tempting to attack a complex problem with complex machinery, as recent research on dis- course parsing has done (see Section 6), we show that a simple model can achieve or surpass state of the art results for discourse relation labeling with a little inspiration from human discourse processing. First, because a decision about whether two EDUs are attached by a semantic relation generally requires reasoning about their contents together, our transformer-based model encodes embeddings of EDU pairs, exploiting a form of message pass- ing simpler than graph neural net models (Wang et al., 2021) while achieving better results. Next, we draw on the fact that for human annotators, the tasks of discourse attachment and relation label- ing are often interdependent: sometimes one sees how to attach two EDUs but only later how to de- termine the relation that links them; sometimes, a clue, e.g. an explicit marker like because, makes the relation clear, and the task is to find the second argument to the relation. Our model exploits this interdependence with a multi-task architecture for attachment prediction and relation labeling. Finally, our model allows for the non-tree-like structures. In Section 2, we provide background on dis- course structure and data sets. In Sections 3 and 4, we describe our model and results. Section 6 presents related work in discourse parsing. 2 Discourse parsing theories and data …

Content truncated to 500 words