Section: New Results

Pairwise coreference models

Participant : Emmanuel Lassalle.

In collaboration with Pascal Denis (Magnet, Inria), we have proposed a new method for significantly improving the performance of pairwise coreference models [34] . Given a set of indicators, our method learns how to best separate types of mention pairs into equivalence classes for which we construct distinct classification models. In effect, our approach finds an optimal fea- ture space (derived from a base feature set and indicator set) for discriminating coreferential mention pairs. Although our approach explores a very large space of possible feature spaces, it remains tractable by exploiting the structure of the hierarchies built from the indicators.

In the framework of decision trees, this method can be seen as a pruning procedure and thus can be combined with different methods for expanding a decision tree. It can also be compared to polynomial kernels, but has the advantage of a lower computational complexity [21] . Our experiments on the CoNLL-2012 Shared Task English datasets (gold mentions) indicate that our method is robust relative to different clustering strategies and evaluation metrics, showing large and consistent improvements over a single pairwise model using the same base features. Our best system obtains a competitive 67.2 of average F1 over MUC, B3, and CEAF which, despite its simplicity, places it above the mean score of other systems on these datasets.