Research On Text Matching Method Based On Deep Learning

Text matching refers to the similarity for measuring two pieces of text at the semantic or intent level.The text matching task belongs to the fundamental research of natural language processing,and its effect has an essential impact on many downstream tasks such as text entailment,automatic question-answer system,and information retrieval.Traditional text matching methods,usually rely on literal features of texts and manually defined rules to measure whether text matching or not,ignore the deep semantic matching features between different texts.Compared with traditional text matching methods,deep learning methods can effectively capture the deep textual semantic features,with excellent model generalization ability,and perform well on the text matching task.This paper focuses on the impact of various text granularities on the textual semantic feature representation,a core task in text matching,and optimizes this representation by integrating semantic features of different granularities to improve the text matching effect.The contributions of this paper are mainly in the following three aspects.(1)Hardly capturing the textual semantic feature information comprehensively,a single sequence encoding model is prone to lose semantic features,which affects the text matching effect.For solving this problem,this paper proposes a text matching method: Single-semantic Feature Fusion."Single semantic" means capturing textual semantic features from a single granularity level,which integrates single semantic feature representations of multi-sequence encodings,and different sequence encodings can capture textual semantic features from various perspectives,which can alleviate the loss of semantic features to a certain extent.Also,this paper proposes a new loss function based on confidence intervals to improve the effect of classifying the model on the instances that are hard to distinguish correctly.The experimental results show that the method can capture richer semantic features and effectively improve the text matching effect.(2)Although the single-semantic feature fusion method in(1)can alleviate the loss of textual semantic features by integrating semantic feature representations from multisequence encodings,while it cannot capture textual semantic features at different granularities simultaneously,which is not comprehensive enough to the acquisition of textual semantic features.In order to solve this problem,this paper proposes a text matching method based on multi-semantic feature fusion."Multi-semantic" means capturing multi-granularity textual semantic features.By considering multi-granularity textual semantic features,the model in this paper can acquire semantic features of different granularities all together,which can further lessen the loss of semantic features.Furthermore,a new loss function is designed to optimize the cross-entropy loss function by using a mean square error(MSE)as a balancing factor.The experimental results show that this method can significantly improve the text matching effect.(3)The multi-semantic feature fusion method in(2)can integrate textual multigranularity semantic features while it does not consider the interaction of textual semantic feature representations between different granularities.A text matching method aimed at solving such issues is proposed by the interactions of multi-semantic features in this paper.It can capture the multi-granularity textual semantic features and deeply explore the interactions of semantic feature representations between multigranularity.This method also further alleviates the loss of semantic features,thus effectively improving the performance of text matching.The experimental results show that it can achieve roughly equal results with BERT,but the corresponding number of parameters is much less than that of BERT.