Coreference Resolution Method Using a Convolutional Neural Network

Abstract

Coreference resolution is one of background tasks in the area of natural language processing. Corresponding state-of-the-art methods are based on machine learning algorithms and consist in detection of regularities between semantic and grammatical features of text entities. In this paper, the method of coreferent pairs detection based on the set of filtering sieves and a convolutional neural network has been suggested. Different current methods to detect coreferent pairs within a text have been analyzed. The set of filtering sieves to find candidates for the formation of coreferent pairs has been implemented according to the features of the Ukrainian language. A multichannel convolutional neural network has been designed. The training of the convolutional neural network has been performed. The effectiveness of the method on the Ukrainian-language corpus has been examined using typical metrics to the coreference resolution task. The results obtained may indicate that the method suggested can be used to solve the coreference resolution task for Ukrainian texts. The method can be adapted according to features of other languages and be applied to corresponding text corpora.