Multi-Perspective Interactive Model for Chinese Sentence Semantic Matching

Chinese sentence semantic matching is a fundamental task in natural language processing, which aims to distinguish whether two Chinese sentences are semantically similar or not. Originated from English semantic matching task, most existing matching methods merely focus on learning the sentence representation from word granularity, but neglect the uniqueness of Chinese characters and the semantic interactions within a sentence on different granularities, and the interactions between sentences. As a result, most existing matching methods on Chinese language only achieve very limited performance improvement. In the paper, we propose a multi-perspective interactive (MPI) model for Chinese sentence semantic matching, which first employs a multi-granularity encoding layer to transform the characters and words in sentences into their embedding representation, then devises a multi-perspective interactive layer to capture the intra-sentence interactions within a sentence but on different granularities and the inter-sentence interactions between sentences. Finally, a prediction layer takes all the captured interactions as input to estimate the matching degree. We also conduct extensive experiments on real-world data set to assess the model performance. The extensive experimental results demonstrate that our proposed model achieves significantly better performance than the compared benchmarks.