Sentence Similarity Computation by Integrating Shallow and Deep Information

Abstract

The calculation of sentence similarity has always been a hot topic in the field of natural language processing. The existing methods of sentence similarity computation usually consider either shallow or deep information of sentences. In this paper, we propose a method to integrating the two types of information. In the model, the shallow information is acquired based on word similarity and sentence length. The deep information is extracted via a parallel convolutional neural network. Then we integrate the two parts linearly. Experiments on SemEval 2007 task 5, similarity evaluation between Chinese and English sentences, show that the performance of our integration model is better than the models which treat the two types of information separately. For parallel convolutional neural network, a performance improvement is evident when shared parameters are assigned.