Discriminative Reranking of Discourse Parses Using Tree Kernels

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2049–2060, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Discriminative Reranking of Discourse Parses Using Tree Kernels Shafiq Joty and Alessandro Moschitti ALT Research Group Qatar Computing Research Institute {sjoty,amoschitti}@qf.org.qa Abstract In this paper, we present a discrimina- tive approach for reranking discourse trees generated by an existing probabilistic dis- course parser. The reranker relies on tree kernels (TKs) to capture the global depen- dencies between discourse units in a tree. In particular, we design new computa- tional structures of discourse trees, which combined with standard TKs, originate novel discourse TKs. The empirical evalu- ation shows that our reranker can improve the state-of-the-art sentence-level parsing accuracy from 79.77% to 82.15%, a rel- ative error reduction of 11.8%, which in turn pushes the state-of-the-art document- level accuracy from 55.8% to 57.3%. 1 Introduction Clauses and sentences in a well-written text are interrelated and exhibit a coherence structure. Rhetorical Structure Theory (RST) (Mann and Thompson, 1988) represents the coherence struc- ture of a text by a labeled tree, called discourse tree (DT) as shown in Figure 1. The leaves cor- respond to contiguous clause-like units called ele- mentary discourse units (EDUs). Adjacent EDUs and larger discourse units are hierarchically con- nected by coherence relations (e.g., ELABORA- TION, CAUSE). Discourse units connected by a re- lation are further distinguished depending on their relative importance: nuclei are the core parts of the relation while satellites are the supportive ones. Conventionally, discourse analysis in RST in- volves two subtasks: (i) discourse segmentation: breaking the text into a sequence of EDUs, and (ii) discourse parsing: linking the discourse units to form a labeled tree. Despite the fact that dis- course analysis is central to many NLP appli- cations, the state-of-the-art document-level dis- course parser (Joty et al., 2013) has an f-score of only 55.83% using manual discourse segmen- tation on the RST Discourse Treebank (RST-DT). Although recent work has proposed rich lin- guistic features (Feng and Hirst, 2012) and pow- erful parsing models (Joty et al., 2012), discourse parsing remains a hard task, partly because these approaches do not consider global features and long range structural dependencies between DT constituents. For example, consider the human- annotated DT (Figure 1a) and the DT generated by the discourse parser of Joty et al. (2013) (Figure 1b) for the same text. The parser makes a mistake in finding the right structure: it considers only e3 as the text to be attributed to e2, where all the text spans from e3 to e6 (linked by CAUSE and ELAB- ORATION) compose the statement to be attributed. Such errors occur because existing systems do not encode long range dependencies between DT con- stituents such as those between e3 and e4−6. Reranking models can make the global struc- tural information available to the system as fol- lows: first, a base parser produces several DT hypotheses; and then a classifier exploits the en- tire information in each hypothesis, …

Content truncated to 500 words