Text Matching Model with Multi-granularity Term Alignment

Abstract

Text matching is one of the fundamental research tasks in the field of natural language processing. It can be applied to a large number of NLP tasks, such as information retrieval, question, and answer systems and text repetition. In this paper, we propose a text-matching model with multi-granularity term alignment (MGTA). The model extracts word information at different granularities through convolutional neural networks and enhances the model effect by aligning the original location features at different word granularities, enabling the model to obtain multiple granularities of information during text matching. We conduct experiments on the Q&A dataset, the text-implication dataset, and the paraphrase recognition dataset, respectively, and compare them with current mainstream models in terms of accuracy, MAP and MRR evaluation metrics, and has fewer parameters, which greatly improves the inference speed.