The utility of parse-derived features for automatic discourse segmentation

Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 488–495, Prague, Czech Republic, June 2007. c⃝2007 Association for Computational Linguistics The utility of parse-derived features for automatic discourse segmentation Seeger Fisher and Brian Roark Center for Spoken Language Understanding, OGI School of Science & Engineering Oregon Health & Science University, Beaverton, Oregon, 97006 USA {fishers,roark}@cslu.ogi.edu Abstract We investigate different feature sets for performing automatic sentence-level dis- course segmentation within a general ma- chine learning approach, including features derived from either finite-state or context- free annotations. We achieve the best re- ported performance on this task, and demon- strate that our SPADE-inspired context-free features are critical to achieving this level of accuracy. This counters recent results sug- gesting that purely finite-state approaches can perform competitively. 1 Introduction Discourse structure annotations have been demon- strated to be of high utility for a number of NLP applications, including automatic text summariza- tion (Marcu, 1998; Marcu, 1999; Cristea et al., 2005), sentence compression (Sporleder and Lap- ata, 2005), natural language generation (Prasad et al., 2005) and question answering (Verberne et al., 2006). These annotations include sentence segmen- tation into discourse units along with the linking of discourse units, both within and across sentence boundaries, into a labeled hierarchical structure. For example, the tree in Figure 1 shows a sentence-level discourse tree for the string “Prices have dropped but remain quite high, according to CEO Smith,” which has three discourse segments, each labeled with ei- ther “Nucleus” or “Satellite” depending on how cen- tral the segment is to the coherence of the text. There are a number of corpora annotated with discourse structure, including the well-known RST Treebank (Carlson et al., 2002); the Discourse GraphBank (Wolf and Gibson, 2005); and the Penn Discourse Treebank (Miltsakaki et al., 2004). While the annotation approaches differ across these cor- pora, the requirement of sentence segmentation into Root ������ HHHHH H Nucleus ���� � HHHH H Nucleus  PPP P Prices have dropped Satellite  PPP P but remain quite high Satellite  PPPP P according to CEO Smith Figure 1: Example Nucleus/Satellite labeled sentence-level discourse tree. sub-sentential discourse units is shared across all ap- proaches. These resources have facilitated research into stochastic models and algorithms for automatic discourse structure annotation in recent years. Using the RST Treebank as training and evalua- tion data, Soricut and Marcu (2003) demonstrated that their automatic sentence-level discourse pars- ing system could achieve near-human levels of ac- curacy, if it was provided with manual segmenta- tions and manual parse trees. Manual segmenta- tion was primarily responsible for this performance boost over their fully automatic system, thus mak- ing the case that automatic discourse segmentation is the primary impediment to high accuracy automatic sentence-level discourse structure annotation. Their models and algorithm – subsequently packaged to- gether into the publicly available SPADE discourse parser1 – make use of the output of the Charniak (2000) parser to derive syntactic indicator features for segmentation and discourse parsing. Sporleder and Lapata (2005) also used …

Content truncated to 500 words