Measurement of sentence similarity based on constituency parsing and dilated convolution

Abstract

Measurement of sentence similarity is widely used in the field of natural language processing, the current mainstream method is based on neural network similarity model. In actual application, the method of neural network has some disadvantages. On the one hand, when sentences are input to the neural network, the problem of semantic loss is caused by the interception and zero-filling operation of the sentence that is too long or too short. On the other hand, ignore the semantic relation between interval words. Thus, this paper proposes a method of measuring sentence similarity based on constituency parsing and dilated convolution. It uses constituency parsing to design rules to reduce unimportant semantic component of long sentences and supplement important semantic component of short sentences. In addition, the receptive fields in sentence dimension and word vector dimension are dilated to capture the semantic association of the two-dimensional interval words. Finally, the method is verified on two data sets.