Data Programming for Learning Discourse Structure

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 640–645 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 640 Data Programming for Learning Discourse Structure Sonia Badene1,2, Kate Thompson1,3, Jean-Pierre Lorr´e2, Nicholas Asher1,3 1IRIT, 2Linagora, 3Universit´e Toulouse III & CNRS {sonia.badene,kate.thompson,nicholas.asher}@irit.fr, {sbadene,jplorre}@linagora.com Abstract This paper investigates the advantages and limits of data programming for the task of learning discourse structure. The data programming paradigm implemented in the Snorkel framework allows a user to label train- ing data using expert-composed heuristics, which are then transformed via the “generative step” into probability distributions of the class labels given the training candidates. These re- sults are later generalized using a discrimina- tive model. Snorkel’s attractive promise to cre- ate a large amount of annotated data from a smaller set of training data by unifying the out- put of a set of heuristics has yet to be used for computationally difficult tasks, such as that of discourse attachment, in which one must de- cide where a given discourse unit attaches to other units in a text in order to form a coher- ent discourse structure. Although approaching this problem using Snorkel requires significant modifications to the structure of the heuristics, we show that weak supervision methods can be more than competitive with classical super- vised learning approaches to the attachment problem. 1 Introduction Discourse structures for texts represent relational semantic structures that convey causal, topical, ar- gumentative relations inter alia or more gener- ally coherence relations. Following (Muller et al., 2012; Li et al., 2014; Morey et al., 2018), we rep- resent them as dependency structures or graphs containing a set of nodes that represent discourse units (DUs), or instances of propositional content, and a set of labelled arcs that represent coherent relations between DUs. For dialogues with mul- tiple interlocutors, extraction of their discourse structures could provide useful semantic informa- tion to the “downstream” models used, for exam- ple, in the production of intelligent meeting man- agers or the analysis of user interactions in on- line fora. However, despite considerable efforts on computational discourse-analysis (Duverle and Prendinger, 2009; Joty et al., 2013; Ji and Eisen- stein, 2014; Surdeanu et al., 2015; Yoshida et al., 2014; Li et al., 2016), we are still a long way from usable discourse models, especially for dialogue. The problem of extracting full discourse structures is difficult: standard supervised models struggle to capture the sparse long distance attachments, even when relatively large annotated corpora are avail- able. In addition, the annotation process is time consuming and often fraught with errors and dis- agreements, even among expert annotators. This motivated us to explore a weak supervision ap- proach, data programming (Ratner et al., 2016), in which we exploit expert linguistic knowledge in a more compact and consistent rule-based form. In our study, we restrict the structure learning problem to predicting edges or attachments be- tween DU pairs in the dependency graph. After training a supervised deep learning algorithm to predict attachments …

Content truncated to 500 words