Constrained Decoding for Text-Level Discourse Parsing

Proceedings of COLING 2012: Technical Papers, pages 1883–1900, COLING 2012, Mumbai, December 2012. Constrained decoding for text-level discourse parsing Philippe Muller1 Ster gos Afantenos1 Pascal Denis2 Nicholas Asher3 (1) IRIT, Université de Toulouse, France (2) Mostrare, INRIA, France (3) IRIT, CNRS, France {stergos.afantenos,muller,asher}@irit.fr, pascal.denis@inria.fr Abstract This paper presents a novel approach to document-based discourse analysis by performing a global A* search over the space of possible structures while optimizing a global criterion over the set of potential coherence relations. Existing approaches to discourse analysis have so far relied on greedy search strategies or restricted themselves to sentence-level discourse parsing. Another advantage of our approach, over other global alternatives (like Maximum Spanning Tree decoding algorithms), is its flexibility in being able to integrate constraints (including linguistically motivated ones like the Right Frontier Constraint). Finally, our paper provides the first discourse parsing system for French; our evaluation is carried out on the Annodis corpus. While using a lot less training data than earlier approaches than previous work on English, our system manages to achieve state-of-the-art results, with F1-scores of 66.2 and 46.8 when compared to unlabeled and labeled reference structures. Keywords: Discourse Structure, Discourse Parsing, Dependency Structures, Constrained Decoding, A*. 1883 1 Introduction Discourse analysis involves the identification of coherence relations between discourse units, which can be either elementary ones—typically clauses or sentences—or complex ones, spanning larger chunks of text. These larger units play similar roles to elementary ones. Coherence relations categorize discourse units in terms of their argumentative, thematic or causal links to other units. Together these relations and the units they relate form the global structure of a discourse. This structure is important as it reflects the thematic organization at various levels of granularity, and constrains semantic interpretations, for instance with respect to anaphora resolution or temporal interpretation. Discourse processing is thus a crucial part of natural language understanding, and it has potentially many applications—opinion detection and classification, question answering, information extraction, recognizing textual entailment, evaluating text coherence, or knowledge extraction to name a few (Stede, 2011; Verberne et al., 2007; Somasundaran et al., 2009; Lin et al., 2011; Gerber et al., 2010) Producing a discourse structure automatically is a complex task, however. Coherence involves syntactic and lexical factors, and relies heavily on the semantic interpretation of its parts. Most existing work tends to focus on restricted aspects of the full problem. The task of building a discourse structure involves three subtasks: (1) identifying discourse units (DUs), (2) “attaching” DUs to one another, and (3) labeling their link with a coherence relation. Of these, the first one is usually considered easiest (see for instance Hernault et al., 2010), the second is arguably the hardest; and as a consequence, researchers have focussed attention on the third one, labelling discourse relations (Lin et al., 2009; Feng and Hirst, 2012; Sporleder and Lascarides, 2008; Wellner et al., 2006; Louis et al., 2010). Research on discourse structure also divides into two orthogonal categories: some researchers limit themselves to intra-sentential discourse structure (Wellner et al., …

Content truncated to 500 words