Modeling Local Coherence: An Entity-Based Approach

Proceedings of the 43rd Annual Meeting of the ACL, pages 141–148, Ann Arbor, June 2005. c⃝2005 Association for Computational Linguistics Modeling Local Coherence: An Entity-based Approach Regina Barzilay Computer Science and Artificial Intelligence Laboratory Massachusetts Institute of Technology regina@csail.mit.edu Mirella Lapata School of Informatics University of Edinburgh mlap@inf.ed.ac.uk Abstract This paper considers the problem of auto- matic assessment of local coherence. We present a novel entity-based representa- tion of discourse which is inspired by Cen- tering Theory and can be computed au- tomatically from raw text. We view co- herence assessment as a ranking learning problem and show that the proposed dis- course representation supports the effec- tive learning of a ranking function. Our experiments demonstrate that the induced model achieves significantly higher ac- curacy than a state-of-the-art coherence model. 1 Introduction A key requirement for any system that produces text is the coherence of its output. Not surprisingly, a variety of coherence theories have been devel- oped over the years (e.g., Mann and Thomson, 1988; Grosz et al. 1995) and their principles have found application in many symbolic text generation sys- tems (e.g., Scott and de Souza, 1990; Kibble and Power, 2004). The ability of these systems to gener- ate high quality text, almost indistinguishable from human writing, makes the incorporation of coher- ence theories in robust large-scale systems partic- ularly appealing. The task is, however, challenging considering that most previous efforts have relied on handcrafted rules, valid only for limited domains, with no guarantee of scalability or portability (Re- iter and Dale, 2000). Furthermore, coherence con- straints are often embedded in complex representa- tions (e.g., Asher and Lascarides, 2003) which are hard to implement in a robust application. This paper focuses on local coherence, which captures text relatedness at the level of sentence-to- sentence transitions, and is essential for generating globally coherent text. The key premise of our work is that the distribution of entities in locally coherent texts exhibits certain regularities. This assumption is not arbitrary — some of these regularities have been recognized in Centering Theory (Grosz et al., 1995) and other entity-based theories of discourse. The algorithm introduced in the paper automat- ically abstracts a text into a set of entity transi- tion sequences, a representation that reflects distri- butional, syntactic, and referential information about discourse entities. We argue that this representation of discourse allows the system to learn the proper- ties of locally coherent texts opportunistically from a given corpus, without recourse to manual annota- tion or a predefined knowledge base. We view coherence assessment as a ranking prob- lem and present an efficiently learnable model that orders alternative renderings of the same informa- tion based on their degree of local coherence. Such a mechanism is particularly appropriate for gener- ation and summarization systems as they can pro- duce multiple text realizations of the same underly- ing content, either by varying parameter values, or by relaxing constraints that control the generation process. A system equipped with a ranking mech- anism, could compare the quality …

Content truncated to 500 words