Easy-first Coreference Resolution

Proceedings of COLING 2012: Technical Papers, pages 2519–2534, COLING 2012, Mumbai, December 2012. Easy-first Coreference Resolution Veselin Sto yanov1 Jason Eisner1,2 (1) HLT-COE, Johns Hopkins University (2) CLSP, Johns Hopkins University {ves,jason}@cs.jhu.edu Abstract We describe an approach to coreference resolution that relies on the intuition that easy decisions should be made early, while harder decisions should be left for later when more information is available. We are inspired by the recent success of the rule-based system of Raghunathan et al. (2010), which relies on the same intuition. Our system, however, automatically learns from training data what constitutes an easy decision. Thus, we can utilize more features, learn more precise weights, and adapt to any dataset for which training data is available. Experiments show that our system outperforms recent state-of-the-art coreference systems including Raghunathan et al.’s system as well as a competitive baseline that uses a pairwise classifier. KEYWORDS: coreference resolution, discourse processing, supervised clustering, greedy ap- proaches. KEYWORDS IN L2: resolución de correferencia. 2519 1 Introduction Coreference resolution has traditionally benefited from machine learning approaches (Soon et al., 2001; Ng and Cardie, 2002; Bengtson and Roth, 2008; Stoyanov et al., 2009). Surprisingly, however, recent work has shown that simple rule-based coreference systems can compete with the state-of-the-art machine-learning-based systems if provided with rich lexical, syntactic, semantic and discourse information (Haghighi and Klein, 2010; Raghunathan et al., 2010). In fact, the rule-based system of Raghunathan et al. (2010) exhibited the top score in the recent CoNLL evaluation (Pradhan et al., 2011). This system’s innovation is to build the coreference clusters incrementally, starting with the most precise rules (dubbed sieves by Raghunathan et al. (2010)) and use the available coreference information to guide the less precise sieves. Coreference resolution is inherently a global task: for example, discovering that Smith and she corefer makes it more probable that Smith corefers with Jane Smith (i.e., a female name) rather than Jason Smith. Furthermore, grouping Jane Smith, she and Jason Smith in the same cluster should be rejected by the clustering algorithm because it results in poor gender and proper name agreement. Yet incorporating such structured information in coreference has proven challenging. There have been several successful attempts to incorporate structured information through joint inference such as Culotta et al. (2007) and Poon and Domingos (2008), but they do not explicitly learn parameters tuned to the inference algorithm used at test time (and the latter relies on a complicated and expensive inference procedures). The difficulty of inferring globally consistent clusterings lies in the fact that there are exponentially many clusterings and producing an “optimal” clustering according to most measures of global coherence is NP-hard. In this context, the approach of Raghunathan et al. (2010) can be seen as a rule-based greedy search for a globally consistent clustering. We propose a coreference resolution approach that like Raghunathan et al. (2010) aims to consider global consistency while performing fast and deterministic greedy search. Similar to Raghunathan et al. (2010), our algorithm operates by making the easy (most …

Content truncated to 500 words