A Constituent-Based Approach to Argument Labeling with Joint Inference in Discourse Parsing

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 68–77, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics A Constituent-Based Approach to Argument Labeling with Joint Inference in Discourse Parsing Fang Kong1∗ Hwee Tou Ng2 Guodong Zhou1 1 School of Computer Science and Technology, Soochow University, China 2 Department of Computer Science, National University of Singapore kongfang@suda.edu.cn nght@comp.nus.edu.sg gdzhou@suda.edu.cn Abstract Discourse parsing is a challenging task and plays a critical role in discourse anal- ysis. In this paper, we focus on label- ing full argument spans of discourse con- nectives in the Penn Discourse Treebank (PDTB). Previous studies cast this task as a linear tagging or subtree extraction problem. In this paper, we propose a novel constituent-based approach to argu- ment labeling, which integrates the ad- vantages of both linear tagging and sub- tree extraction. In particular, the pro- posed approach unifies intra- and inter- sentence cases by treating the immediate- ly preceding sentence as a special con- stituent. Besides, a joint inference mech- anism is introduced to incorporate glob- al information across arguments into our constituent-based approach via integer lin- ear programming. Evaluation on PDT- B shows significant performance improve- ments of our constituent-based approach over the best state-of-the-art system. It al- so shows the effectiveness of our joint in- ference mechanism in modeling global in- formation across arguments. 1 Introduction Discourse parsing determines the internal struc- ture of a text and identifies the discourse rela- tions between its text units. It has attracted in- creasing attention in recent years due to its impor- tance in text understanding, especially since the release of the Penn Discourse Treebank (PDTB) corpus (Prasad et al., 2008), which adds a layer of discourse annotations on top of the Penn Treebank ∗The research reported in this paper was carried out while Fang Kong was a research fellow at the National University of Singapore. (PTB) corpus (Marcus et al., 1993). As the largest available discourse corpus, the PDTB corpus has become the defacto benchmark in recent studies on discourse parsing. Compared to connective identification and dis- course relation classification in discourse parsing, the task of labeling full argument spans of dis- course connectives is much harder and thus more challenging. For connective identification, Lin et al. (2014) achieved the performance of 95.76% and 93.62% in F-measure using gold-standard and automatic parse trees, respectively. For discourse relation classification, Lin et al. (2014) achieved the performance of 86.77% in F-measure on clas- sifying discourse relations into 16 level 2 types. However, for argument labeling, Lin et al. (2014) only achieved the performance of 53.85% in F- measure using gold-standard parse trees and con- nectives, much lower than the inter-annotation a- greement of 90.20% (Miltsakaki et al., 2004). In this paper, we focus on argument labeling in the PDTB corpus. In particular, we propose a nov- el constituent-based approach to argument label- ing which views constituents as candidate argu- ments. Besides, our approach unifies intra- and inter-sentence cases by treating the …

Content truncated to 500 words