Window-Based Neural Tagging for Shallow Discourse Argument Labeling

Proceedings of the 23rd Conference on Computational Natural Language Learning, pages 768–777 Hong Kong, China, November 3-4, 2019. c⃝2019 Association for Computational Linguistics 768 Window-Based Neural Tagging for Shallow Discourse Argument Labeling René Knaebel, Manfred Stede Applied Computational Linguistics Department of Linguistics University of Potsdam Germany {rknaebel,stede}@uni-potsdam.de Sebastian Stober Artificial Intelligence Lab Faculty of Computer Science Otto von Guericke University Magdeburg Germany stober@ovgu.de Abstract This paper describes a novel approach for the task of end-to-end argument labeling in shal- low discourse parsing. Our method describes a decomposition of the overall labeling task into subtasks and a general distance-based aggrega- tion procedure. For learning these subtasks, we train a recurrent neural network and gradually replace existing components of our baseline by our model. The model is trained and evalu- ated on the Penn Discourse Treebank 2 corpus. While it is not as good as knowledge-intensive approaches, it clearly outperforms other mod- els that are also trained without additional lin- guistic features. 1 Introduction Shallow discourse parsing (SDP) is a challeng- ing problem in NLP with the aim to identify lo- cal coherence relations in text. Discourse rela- tions are used in text to connect individual seg- ments of text logically. The Penn Discourse Tree- bank (PDTB) (Prasad et al., 2008a) adopts a non- hierarchical view on discourse relations. As an ex- ample from the PDTB, the sentence: • We would stop index arbitrage when the mar- ket is under stress. contains an explicit discourse relation which is sig- naled through the underlined connective (Conn) and further consists of two arguments (Arg1 in italics and Arg2 in bold). In addition, a sense is assigned to a relation, such as Condition. Discourse analysis, however, does not work on a sentence-level, but takes full documents into ac- count. Often, short paragraphs suffice to show the challenge in extracting overlapping relations. To il- lustrate the problem, we split the paragraphs shown in Figure 1 into small text chunks. Though our im- plementation works on the level of individual to- If you think you have stress-related problems on the job, there’s good news and bad news. You’re probably right, and you aren’t alone. ... Even the courts are beginning to recognize the link between jobs and stress-related disorders in compensation cases, ac- cording to a survey by the National Council on Compensation Insurance. But although 56% of the respondents in the study indicated that mental-health problems were fairly pervasive in the workplace, there is still a social stigma associated with people seeking help. Figure 1: Excerpt from text WSJ 1582 (PDTB corpus) Text Chunk R1 R2 R3 If Conn Arg1 you think you have stress-related problems on the job, Arg2 Arg1 there’s good news and bad news. Arg1 Arg1 You’re probably right, Arg2 Arg1 and Arg2 Conn you aren’t alone Arg2 Arg2 Table 1: First sample paragraph split into text chunks. kens, we here use these chunks to highlight chal- lenges in discourse argument labeling. The chunks are delimited such that no smaller part would play …

Content truncated to 500 words