Understanding the Value of Features for Coreference Resolution

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 294–303, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Understanding the Value of Features for Coreference Resolution Eric Bengtson Dan Roth Department of Computer Science University of Illinois Urbana, IL 61801 {ebengt2,danr}@illinois.edu Abstract In recent years there has been substantial work on the important problem of coreference res- olution, most of which has concentrated on the development of new models and algo- rithmic techniques. These works often show that complex models improve over a weak pairwise baseline. However, less attention has been given to the importance of selecting strong features to support learning a corefer- ence model. This paper describes a rather simple pair- wise classification model for coreference res- olution, developed with a well-designed set of features. We show that this produces a state-of-the-art system that outperforms sys- tems built with complex models. We suggest that our system can be used as a baseline for the development of more complex models – which may have less impact when a more ro- bust set of features is used. The paper also presents an ablation study and discusses the relative contributions of various features. 1 Introduction Coreference resolution is the task of grouping all the mentions of entities1 in a document into equivalence classes so that all the mentions in a given class refer to the same discourse entity. For example, given the sentence (where the head noun of each mention is subscripted) 1We follow the ACE (NIST, 2004) terminology: A noun phrase referring to a discourse entity is called a mention, and an equivalence class is called an entity. An American1 official2 announced that American1 President3 Bill Clinton3 met his3 Russian4 counterpart5, Vladimir Putin5, today. the task is to group the mentions so that those refer- ring to the same entity are placed together into an equivalence class. Many NLP tasks detect attributes, actions, and relations between discourse entities. In order to discover all information about a given entity, tex- tual mentions of that entity must be grouped to- gether. Thus coreference is an important prerequi- site to such tasks as textual entailment and informa- tion extraction, among others. Although coreference resolution has received much attention, that attention has not focused on the relative impact of high-quality features. Thus, while many structural innovations in the modeling ap- proach have been made, those innovations have gen- erally been tested on systems with features whose strength has not been established, and compared to weak pairwise baselines. As a result, it is possible that some modeling innovations may have less im- pact or applicability when applied to a stronger base- line system. This paper introduces a rather simple but state- of-the-art system, which we intend to be used as a strong baseline to evaluate the impact of structural innovations. To this end, we combine an effective coreference classification model with a strong set of features, and present an ablation study to show the relative impact of a variety of features. …

Content truncated to 500 words