Limited memory incremental coreference resolution

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 2129–2139, Dublin, Ireland, August 23-29 2014. Limited memory incremental coreference resolution Kellie Webster and James R. Curran e -lab, School of Information Technologies University of Sydney NSW 2006, Australia {kellie.webster,james.r.curran}@sydney.edu.au Abstract We propose an algorithm for coreference resolution based on analogy with shift-reduce pars- ing. By reconceptualising the task in this way, we unite ranking- and cluster-based approaches to coreference resolution, which have until now been largely orthogonal. Additionally, our framework naturally lends itself to rich discourse modelling, which we use to define a series of psycholinguistically motivated features. We achieve CoNLL scores of 63.33 and 62.91 on the CoNLL-2012 DEV and TEST splits of the OntoNotes 5 corpus, beating the publicly available state of the art systems. These results are also competitive with the best reported research systems despite our system having low memory requirements and a simpler model. 1 Introduction Coreference resolution is the task of partitioning mentions in a document, usually noun phrases, into clusters which correspond to their real world referents. It is typically approached as a classification task between mentions; given a set of mentions, systems predict the likelihood of their being coreferential with one another and translate these scores into a clustering in a decoding phase. The task has received considerable research attention due to its importance for downstream inference in tasks such as named entity linking and relation extraction. While simple, local models of coreference have established a reasonable baseline, encoding global consistency requirements remains a challenge since their complete representation is computationally intractable. Two promising but orthogonal direc- tions addressing the need for global consistency measures are ranking-based decoding (Ng and Cardie, 2002; Denis et al., 2007; Fernandes et al., 2012; Durrett and Klein, 2013; Chang et al., 2013) and cluster-based modelling (Rahman and Ng, 2009; Raghunathan et al., 2010; Lee et al., 2011; Klenner and Tuggener, 2011). However, among current systems, decoding strategies are increasingly complex and cluster-based models do not fully leverage psycholinguistic cues such as reading order. The primary contribution of our work is a reconceptualisation of the coreference task by analogy with the shift-reduce parsing algorithm. This reconceptualisation allows us to capitilise on both ranking- and cluster-based approaches and our system, LIMERIC, outperforms systems using either approach in isolation. We go beyond the shift-reduce algorithm by interpreting our stack of partially formed clusters as a reader’s mental status while reading. This allows us to introduce a series of rich discourse features which capture antecedent competition and cognitive accessiblity via a cluster’s position in the stack. Our system is simple and efficient, using maximum-margin averaged perceptron classification and optional beam-search decoding during inference. Despite requiring only a limited amount of memory, our system achieves the competitive CoNLL scores of 63.33 and 62.91 on the CoNLL-2012 DEV and TEST splits of the OntoNotes 5 corpus (Pradhan et al., 2012). We argue that this is due to its more faithful representation of cognitive processing and that extending …

Content truncated to 500 words