A State-of-the-Art Mention-Pair Model for Coreference Resolution

Proceedings of the Fourth Joint Conference on Lexical and Computational Semantics (*SEM 2015), pages 289–298, Denver, Colorado, June 4–5, 2015. A State-of-the-Art Mention-Pair Model for Coreference Resolution Olga Uryupina1 and Alessandro Moschitti2,1 1Department of Information Engineering and Computer Science, University of Trento, 2Qatar Computing Research Institute uryupina@gmail.com, amoschitti@gmail.com Abstract Most recent studies on coreference resolu- tion advocate accurate yet relatively com- plex models, relying on, for example, entity- mention or graph-based representations. As it has been convincingly demonstrated at the recent CoNLL 2012 shared task, such algo- rithms considerably outperform popular basic approaches, in particular mention-pair mod- els. This study advocates a novel approach that keeps the simplicity of a mention-pair framework, while showing state-of-the-art re- sults. Apart from being very efficient and straightforward to implement, our model fa- cilitates experimental work on the pairwise classifier, in particular on feature engineering. The proposed model achieves the performance level of up to 61.82% (MELA F, v4 scorer) on the CoNLL test data, on par with complex state-of-the-art systems. 1 Introduction The mention-pair model, as proposed by Soon et al. (2001) has been used for over a decade now. It combines a simple classifier trained to discrimi- nate between coreferent and not-coreferent pairs of mentions (“links”) with fast heuristic procedures for merging the classifier’s decisions at the decoding stage. Several decoding heuristics have been advo- cated in the literature, the most commonly used ones including first-link (Soon et al., 2001) and best-link (Ng and Cardie, 2002). Most state-of-the-art algorithms for coreference resolution, on the contrary, rely on complex mod- eling, ranging from entity-ranking to structural per- ceptron and other graph-based approaches (for an overview of state-of-the-art coreference resolvers, see (Ng, 2010; Pradhan et al., 2012)). Such algo- rithms show a clearly superior performance: thus, at the CoNLL-2012 shared task, the best-performing (Soon et al., 2001)-style system loses around 8% to the winning algorithm. However, more traditional mention-pair ap- proaches still have some important advantages. Thus, a mention-pair model is easy to implement and allows for fast prototyping. It relies on a sim- ple binary classifier making it very fast to train com- pared to state-of-the-art models that are based on complex structural representations (Fernandes et al., 2012; Bj¨orkelund and Kuhn, 2014). This efficiency at the training step allows for straightforward au- tomatic parameter optimization. Most importantly, mention-pair models can be useful for understand- ing low-level system behavior, and, in particular, for feature engineering. This can in turn help improve more complex models, since many of them rely on mention-pairs as their basic building blocks. In this paper, we advocate a new easy-first mention-pair algorithm (EFMP): while it is based solely on pairs of mentions and does not attempt any global inference, it benefits from the decision propagation strategy to create a coreference parti- tion. Augmented with the sieve-style prefiltering, the system achieves a performance level comparable to the state of the art. The contribution of this paper is two-fold. First, we propose a novel decoding approach that com- bines predictions of the …

Content truncated to 500 words