Sentences Similarity Based on Deep Structured Semantic Model and Semantic Role Labeling*

摘要

This paper proposes a method of sentences similarity based on deep semantic model and semantic role labeling. Deep semantic model is to learn word representation in a semantic space using deep learning technology, then we use generated vectors to calculate sentences similarity. At the same time, we label sentences with semantic roles and calculate similarity based on semantic roles. Finally, those two kinds of sentences similarity are linearly combined as final sentences similarity. In this paper, experiments are carried out on the SemEva1-2017 and SemEva1-2016 task 1, and Pearson correlation coefficients reach 85% and 63% respectively. Experimental results show that our proposed method outperforms existing methods demonstrating the effectiveness of our approach.