Research On Multi-granular Text Matching Method Based On Deep Learning

The main purpose of text matching research is to calculate the similarity between two paragraphs of text.Many tasks can be abstracted into text matching problems to a certain extent,such as: information search,automatic answer,machine translation,dialogue system,paraphrase recognition and so on.The current research mainly uses deep learning methods and deep neural networks to learn the semantic information contained in two paragraphs of text in order to achieve the purpose of accurate matching.Conventional text matching methods are roughly divided into representation-based text matching models and interaction-based text matching models.Since the representationbased text matching model tends to lose its semantic focus,and the interaction-based text matching model ignores global information,and there is different granular phrase information in the text,the single-word information matching adopted by the interactive text matching model may be important matching information is missing.In order to address the above problems,this thesis proposes a fusion model using multi-granularity information.The model combines the two text matching models through interactive attention and represented attention.It not only makes full use of the advantages of the interaction model and the represented model,but also complements the two types of models.And then,by extracting multiple different levels of granular information in the text,the model can not only grasp the local important information,but also obtain the global semantic information.The main research content is as follows:1.FMMI model: This model first uses interactive attention to integrate two different models,so that it can show full advantages of the two types of different models.At the same time,in order to obtain the multi-granularity information existing in the text,a framework of multi-granularity information is designed.This framework extracts granular information of different phrase sizes in the text at a horizontal level through a convolutional neural network to obtain the semantics of multiple granularities contained in the text.It solves the problem that traditional models lose their semantic focus and ignore global information in the process of text matching.Experimental results on three different sets of text matching tasks show that the proposed model can effectively focus on different levels of text by extracting multiple granularity phrase information in the text and fusing interactive attention with represented attention.The three evaluation indicators of NDCG@3,NDCG@5,and MAP have been significantly improved.2.FMHG model: Since the learning representation process of the neural network is from fine-grained to coarse-grained,the output of the network only retains the final coarse-grained information,and the original fine-grained information is lost.The above model(FMMI)can just extract phrase information at the same level.Therefore a text matching fusion model FMHG that combines with hierarchical granularity information extraction is proposed in this thesis.Based on the interactive model,the new model changes the later level granularity extraction method of the model and introduces a hierarchical granularity extraction idea which mainly uses the representation on ability of deep neural networks to make the original tensor information pass through networks of different depths to obtain coarse-grained information and fine-grained information under different representation levels in the network.Experimental results on three different sets of text matching tasks show that the proposed model can effectively focus on different levels of text information by extracting the hierarchical granularity information in the text.The experimental effect is significantly improved compared with the cases without granular information.It also has a improvement on the SNLI data set and the semeval2016-task3 dataset.