Recognizing Textual Parallelisms with Edit Distance and Similarity Degree

Recognizing Textual Parallelisms with edit distance and similarity degree Marie Gu´egan and Nicolas Hernandez LIMSI-CNRS Universit´e de Paris-Sud, France guegan@aist.enst.fr | hernandez@limsi.fr Abstract Detection of discourse structure is crucial in many text-based applications. This pa- per presents an original framework for de- scribing textual parallelism which allows us to generalize various discourse phe- nomena and to propose a unique method to recognize them. With this prospect, we discuss several methods in order to iden- tify the most appropriate one for the prob- lem, and evaluate them based on a manu- ally annotated corpus. 1 Introduction Detection of discourse structure is crucial in many text-based applications such as Information Re- trieval, Question-Answering, Text Browsing, etc. Thanks to a discourse structure one can precisely point out an information, provide it a local context, situate it globally, link it to others. The context of our research is to improve au- tomatic discourse analysis. A key feature of the most popular discourse theories (RST (Mann and Thompson, 1987), SDRT (Asher, 1993), etc.) is the distinction between two sorts of discourse re- lations or rhetorical functions: the subordinating and the coordinating relations (some parts of a text play a subordinate role relative to other parts, while some others have equal importance). In this paper, we focus our attention on a dis- course feature we assume supporting coordination relations, namely the Textual Parallelism. Based on psycholinguistics studies (Dubey et al., 2005), our intuition is that similarities concerning the sur- face, the content and the structure of textual units can be a way for authors to explicit their intention to consider these units with the same rhetorical im- portance. Parallelism can be encountered in many specific discourse structures such as continuity in infor- mation structure (Kruijff-Korbayov´a and Kruijff, 1996), frame structures (Charolles, 1997), VP el- lipses (Hobbs and Kehler, 1997), headings (Sum- mers, 1998), enumerations (Luc et al., 1999), etc. These phenomena are usually treated mostly inde- pendently within individual systems with ad-hoc resource developments. In this work, we argue that, depending on de- scription granularity we can proceed, computing syntagmatic (succession axis of linguistic units) and paradigmatic (substitution axis) similarities between units can allow us to generically handle such discourse structural phenomena. Section 2 introduces the discourse parallelism phenomenon. Section 3 develops three methods we implemented to detect it: a similarity degree measure, a string editing distance (Wagner and Fischer, 1974) and a tree editing distance1 (Zhang and Shasha, 1989). Section 4 discusses and evaluates these methods and their relevance. The final section reviews re- lated work. 2 Textual parallelism Our notion of parallelism is based on similarities between syntagmatic and paradigmatic represen- tations of (constituents of) textual units. These similarities concern various dimensions from shal- low to deeper description: layout, typography, morphology, lexicon, syntax, and semantics. This account is not limited to the semantic dimension as defined by (Hobbs and Kehler, 1997) who con- sider text fragments as parallel if the same predi- cate can be inferred from them with coreferential or similar pairs of …

Content truncated to 500 words