Lightweight Text Matching Method with Rich Features

Text matching is one of the research hotspots in Natural Language Processing (NLP). The study of text matching is of great practical importance for applications such as text de-duplication, web retrieval, and question answering systems. A lightweight text matching method with rich features is proposed for the problem of large number of model parameters and low efficiency of text matching tasks in natural language processing. The whole model architecture is based on Siamese neural networks with shared parameters. Furthermore, the method utilizes an improved residual Network and attention mechanism for the extraction and alignment of vector representations. Only three key features for alignment operations are retained. In addition, an averaging operation is added to the fusion layer to provide vector representations with rich information for the prediction layer. Experimental results on the paraphrase identification dataset and two natural language inference datasets show that the proposed approach not only effectively reduces the number of parameters compared with existing models but also ensures good text matching performance. Experiments demonstrate that this method can be used in general text matching tasks.