A Decision-Based Approach to Rhetorical Parsing

A Decision-Based Approach to Rhetorical Parsing Daniel Marcu Information Sciences Institute and Department of Computer Science University of Southern California 4676 Admiralty Way, Suite 1001 Marina del Rey, CA 90292-6601 marcu @ isi. edu Abstract We present a shift-reduce rhetorical parsing algo- rithm that learns to construct rhetorical structures of texts from a corpus of discourse-parse action se- quences. The algorithm exploits robust lexical, syn- tactic, and semantic knowledge sources. I Introduction The application of decision-based learning tech- niques over rich sets of linguistic features has improved significantly the coverage and perfor- mance of syntactic (and to various degrees seman- tic) parsers (Simmons and Yu, 1992; Magerman, 1995; Hermjakob and Mooney, 1997). In this pa- per, we apply a similar paradigm to developing a rhetorical parser that derives the discourse structure of unrestricted texts. Crucial to our approach is the reliance on a cor- pus of 90 texts which were manually annotated with discourse trees and the adoption of a shift-reduce parsing model that is well-suited for learning. Both the corpus and the parsing model are used to gener- ate learning cases of how texts should be partitioned into elementary discourse units and how discourse units and segments should be assembled into dis- course trees. 2 The Corpus We used a corpus of 90 rhetorical structure trees, which were built manually using rhetorical rela- tions that were defined informally in the style of Mann and Thompson (1988): 30 trees were built for short personal news stories from the MUC7 co- reference corpus (Hirschman and Chinchor, 1997); 30 trees for scientific texts from the Brown corpus; and 30 trees for editorials from the Wall Street Jour- nal (WSJ). The average number of words for each text was 405 in the MUC corpus, 2029 in the Brown corpus, and 878 in the WSJ corpus. Each MUC text 365 was tagged by three annotators; each Brown and WSJ text was tagged by two annotators. The rhetorical structure assigned to each text is a (possibly non-binary) tree whose leaves correspond to elementary discourse units (edu)s, and whose in- ternal nodes correspond to contiguous text spans. Each internal node is characterized by a rhetori- cal relation, such as ELABORATION and CONTRAST. Each relation holds between two non-overlapping text spans called NUCLEUS and SATELLITE. (There are a few exceptions to this rule: some relations, such as SEQUENCE and CONTRAST, are multinu- clear.) The distinction between nuclei and satellites comes from the empirical observation that the nu- cleus expresses what is more essential to the writer's purpose than the satellite. Each node in the tree is also characterized by a promotion set that denotes the units that are important in the corresponding subtree. The promotion sets of leaf nodes are the leaves themselves. The promotion sets of internal nodes are given by the union of the promotion sets of the immediate nuclei nodes. Edus are defined functionally as clauses or clause-like units that are unequivocally the NU- CLEUS or SATELLITE of a rhetorical relation that holds between …

Content truncated to 500 words