An Incremental Entity-Mention Model for Coreference Resolution with Restrictive Antecedent Accessibility

Proceedings of Recent Advances in Natural Language Processing, pages 178–185, Hissar, Bulgaria, 12-14 September 2011. An Incremental Entity-Mention Model for Coreference Resolution with Restrictive Antecedent Accessibility Manfred Klenner University of Zurich Institute of Computational Linguistics klenner@cl.uzh.ch Don Tuggener University of Zurich Institute of Computational Linguistics tuggener@cl.uzh.ch Abstract We introduce an incremental entity- mention model for coreference resolution. Our experiments show that it is superior to a non-incremental version in the same en- vironment. The benefits of an incremental architecture are: a reduction of the num- ber of candidate pairs, a means to over- come the problem of underspecified items in pairwise classification and the natural integration of global constraints such as transitivity. Additionally, we have de- fined a simple salience measure that - cou- pled with the incremental model - proved to establish a challenging baseline which seems to be on par with machine learn- ing based systems of the 2010’s SemEval shared task. 1 Introduction With notable exceptions (Luo et al., 2004; Yang et al., 2004; Daume III and Marcu, 2005; Cu- lotta et al., 2007; Rahman and Ng, 2009; Cai and Strube, 2010; Raghunathan et al., 2010) super- vised approaches to coreference resolution are of- ten realised by pairwise classification of anaphor- antecedent candidates. A popular and often reim- plemented approach is presented in (Soon et al., 2001). As recently discussed in (Ng, 2010), the so called mention-pair model suffers from several design flaws which originate from the locally con- fined perspective of the model: • Generation of (transitively) redundant pairs, as the formation of coreference sets (corefer- ence clustering) is done after pairwise classi- fication • Skewed training sets based on pair generation mechanics which lead to classifiers biased to- wards negative classification • No means to enforce global constraints such as transitivity • Underspecification of antecedent candidates Mention-pair systems operate in a non- incremental mode, i.e. all pairs are classified prior to the construction of the coreference sets. A clustering step is needed where, additionally, inconsistencies (e.g. transitively incompatible pairs) can be removed. This often is realised as an optimisation step, where scores derived from pairwise classification are used as weights in a decision taking process that incorporates linguis- tic constraints, e.g. (Finkel and Manning, 2008). Although this overcomes the limitations of the strictly local perspective of pairwise classifiers, it still suffers from the problem of unbalanced data (much more negative than positive examples are generated). The large number of candidate pairs, in general, is a problem, e.g. (Wunsch et al., 2009). These problems can be remedied by an in- cremental entity-mention model, where candidate pairs are evaluated on the basis of emerging coref- erence sets. The amount of candidate pairs is re- duced, since only one (virtual prototype) example of each coreference set needs to be compared to a new anaphor candidate1. Moreover, the prob- lem of inconsistent decisions vanishes, since the virtual prototype of a coreference set bears all the known morphological and semantic information of the elements of the set. If an anaphor …

Content truncated to 500 words