Learning Structured Perceptrons for Coreference Resolution with Latent Antecedents and Non-local Features

Abstract

We investigate different ways of learning structured perceptron models for coreference resolution when using non-local features and beam search. Our experimental results indicate that standard techniques such as early updates or Learning as Search Optimization (LaSO) perform worse than a greedy baseline that only uses local features. By modifying LaSO to delay updates until the end of each instance we obtain significant improvements over the baseline. Our model obtains the best results to date on recent shared task data for Arabic, Chinese, and English.

Prune-and-Score: Learning for Greedy Coreference Resolution

“…Comparison to State-of-the-Art. Table 4 shows the results of our Prune-and-Score approach compared with the following state-of-theart coreference resolution approaches: HOTCoref system (Björkelund and Kuhn, 2014); Berkeley system with the FINAL feature set ; CPL 3 M system ; Stanford system (Lee et al, 2013); Easy-first system (Stoyanov and Eisner, 2012);andFernandes et al, 2012 (Fernandes et al, 2012). Only Scoring is the special case of our Prune-andScore approach where we employ only the scoring function.…”

“…Therefore, other coreference resolution systems Björkelund and Kuhn, 2014) can also benefit from this idea. One way to further improve the peformance of our approach is to perform a search in the Limited Discrepancy Search (LDS) space (Doppa et al, 2014b) using the learned functions.…”

“…They have shown great success and are very efficient. Indeed, several of the approaches that have achieved state-of-the-art results on OntoNotes fall under this category Björkelund and Kuhn, 2014). However, their efficiency requirement leads to a highly nonrealizable learning problem.…”

“…The pruning constraints can take different forms (e.g., classifiers, decisionlist, or ranking functions) depending on the search architecture. Therefore, other coreference resolution systems Björkelund and Kuhn, 2014) can also benefit from this idea. While our basic idea of twolevel selection might appear similar to the coarseto-fine inference architectures (Felzenszwalb and McAllester, 2007;Weiss and Taskar, 2010), the details differ significantly.…”

Entity-Centric Coreference Resolution with Model Stacking

“…In Table 3 we compare the results of our system with the following state-of-the-art approaches: the JOINT and INDEP models of the Berkeley system (Durrett and Klein, 2014) (the JOINT model jointly does NER and entity linking along with coreference); the Prune-and-Score system (Ma et al, 2014); the HOTCoref system (Björkelund and Kuhn, 2014); the CPL 3 M sytem (Chang et al, 2013); and Fernandes et al We use the full entitycentric clustering algorithm drawing upon scores from both pairwise models. We do not make use of agreement features, as these did not increase accuracy and complicate the system.…”