A Convolutional Deep Neural Network for Coreference Resolution via Modeling Hierarchical Features

Coreference resolution is a major task of natural language processing NLP identifying which noun phrases or mentions refer to the same real-world entity or concept. The state-of-the-art methods applied to coreference resolution are mainly based on statistical machine learning, and their performance strongly depends on the quality of the extracted features. The extracted features are usually shallow features by artificial selection, which leads to the loss of unknown useful deep semantic information and becomes an obstacle for improving system performance. We explored a convolutional deep neural network CDNN to extract discourse level features automatically. Our method utilized all of the word tokens as input without complicated pre-processing. To begin with, the word tokens were transformed to vectors by looking up word embeddings. Secondly, mention-pair level features were extracted according to the given mentions. In the meanwhile, distance features were computed easily. Moreover, discourse level features were learned using a convolutional approach. Finally, these features were fed into a softmax classifier to predict the equivalence between two marked mentions. The experimental results demonstrate that our approach obtains a competitive score of average F1 over MUC, B3, and CEAF, which places it above the mean score of other systems on the dataset of CoNLL-2012 Shared Task.