Learning Global Features for Coreference Resolution

Abstract

There is compelling evidence that coreference prediction would benefit from modeling global information about entity-clusters. Yet, state-of-the-art performance can be achieved with systems treating each mention prediction independently, which we attribute to the inherent difficulty of crafting informative clusterlevel features. We instead propose to use recurrent neural networks (RNNs) to learn latent, global representations of entity clusters directly from their mentions. We show that such representations are especially useful for the prediction of pronominal mentions, and can be incorporated into an end-to-end coreference system that outperforms the state of the art without requiring any additional search.

Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017)

“…We would like to exploit the pragmatic and socio-linguistic context to better handle code-switching. 386 model of Wiseman et al (2015b), which on its own outperforms the original one and achieves similar performance to its global version (Wiseman et al, 2016). Importantly when we use the introduced differentiable relaxations in training, we observe a substantial gain in performance over our probabilistic baseline.…”

“…Most importantly, it yields the second-best results reported in the literature, outperforming the system of Wiseman et al (2016) by 0.26 percentage points.…”

“…Comparison between the State-of-the-Art When trained and evaluated on our dataset, both the Stanford (Clark and Manning, 2016) and the Harvard (Wiseman et al, 2016) systems give comparable results to their performance on the CoNLL'12 dataset. 7 The Stanford system using its pre-trained model gives the µ scores of 47.67% and 64.14% for the episode and scene-level respectively, which signifies the importance of the in-domain training data.…”

“…All these systems represent stateof-the-art research in the field. Note that we do not include the very latest deep learning based approaches (Wiseman et al, 2016;Clark and Manning, 2016) to allow for a fair comparison: since, as we have seen in the experiments above, the collaborative partitioning algorithm consistently improves over individual ensemble components, integrating the very best systems would be a trivial but not very informative way of advancing the state of the art. Table 8 shows the performance level of each of these systems on the English portion of the CoNLL-2012 dataset, individually and of the collaborative ensemble.…”

“…This task is similar to coreference resolution in the sense that it groups mentions into entities, but distinct because it requires the identification of mention groups as real entities. Furthermore, even if it can be tackled as a coreference resolution task, only a few coreference resolution systems are designed to handle dialogues well (Rocha, 1999;Niraula et al, 2014) although several state-of-the-art systems have been proposed for the general domain (Peng et al, 2015;Clark and Manning, 2016;Wiseman et al, 2016).…”