Multi-perspective Text Matching Algorithm Based on Multi-granularity Feature Convolution

The core of Chinese text matching task lies in mining the deep semantic information inside the text, exploring the semantic similarities and differences between different texts, and then analyzing the semantic similarity between two texts to be matched. To address the problems of single text granularity feature, insufficient capture of potential semantic information at multiple granularities and weak interactive matching of coding features in Chinese text matching, we propose a multi-perspective text matching model based on multi-granularity features convolution (MpmMfc). The model first extracts characters, words and associated phrases by multi-pattern partitioning and performs initial encoding, then uses a two-way gate loop control unit to initially extract the contextual semantic information in the encoding. Then a multi-grain size high-dimensional encoding matrix is constructed and convolutional neural network is used to capture the granularity features to improve the characterization of multi-granularity semantic information. Finally, the multi-granularity convolutional matrices of two texts are cross-cosine matched by multi-perspective matching patterns to enhance the interaction strength of multi-granularity feature information. The results achieved in the experiments of LCQMC, a spoken expression class dataset, and BQ, a financial class dataset, are better than the currently available non-BERT type text matching models.