Measuring text similarity based on structure and word embedding

Abstract

The problem of finding the similarity between natural language sentences is crucial for many applications in Natural Language Processing (NLP). An accurate calculation of similarity between sentences is highly needed. Many approaches depend on word-to-word similarity to measure sentence similarity. This paper proposes a new approach to improve the accuracy of the sentence similarity calculation. The proposed approach combines different similarity measures in the calculation of sentence similarity. In addition to traditional word-to-word similarity measure, the proposed approach exploits sentence semantic structure. Discourse representation structure (DRS) which is a semantic representation for natural sentences is generated and used to calculated structure similarity. Furthermore, word order similarity is measured to consider the order of words in sentences. Experiments show that exploiting structural information achieves good results. Moreover, the proposed method outperforms the current approaches on a standard benchmark dataset achieving 0.8813 Pearson correlation with human similarity.