Paraphrase Identification Based on Interpretable Mechanism

Paraphrase identification is an important branch of natural language understanding. This paper proposes a vector representation method integrating explanatory text by adding the semantic features of explanatory text to the word vector representation, the vector representation integrates the common interpretation information of words outside the sentence, so as to enrich the semantics of vector representation. This paper obtains explanatory texts from the corpus of Modern Chinese Dictionary. Using a variety of neural network structures on the LCQMC, this work obtained more than 1% performance improvement with fusing explanatory text. The experimental results show that the vector representation of fused explanatory text is more suitable for the task of paraphrase recognition.