Coreferent Mention Detection using Deep Learning

A mention may or may not be coreferred elsewhere in the document. Identifying those mentions that are corefered (called coreferents) is an important step in many NLP tasks, like coreference resolution. To classify a mention as singleton or coreferent using just one sentence is a challenging problem, but previous work suggests that there are cues in a sentence which can be used to predict if a mention it contains is corefered elsewhere in the document. We approach this problem in two different ways. First, we try to a classify a mention candidate extracted by rules-based methods as a coreferent or singleton using various hand-crafted features from literature and a Feed Forward Neural Network (FFNN). Second, we approach this problem as identifying coreferent mention boundaries in input sequences using Recurrent Neural Networks. This is a two step process where we first detect mention heads, and then the mention boundaries. Both these parts are trained and evaluated independently. The second approach removes our dependency on an upstream rule-based mention extractor and hand-crafted features for classification, some of which are expensive to compute. Our hypothesis is that the second approach would be able to learn those hand-crafted features (and more) automatically and perform better at the task. We observed that both of our approaches outperformed the baseline logistic regression model which uses all the hand-crafted features of the first approach, showing that deep neural networks can be used effectively for coreferent mention detection. We also observed that the presence of hand crafted features from literature helps and the first approach (FFNN) outperformed the second (RNN) by approximately 3 F1 points. We also compare our results to published results for the coreferent mention detection task.