Deep Sentence Interactive Matching Model Based On Multi-perspective Feature Fusion

Sentence matching is a key problem in natural language understanding,so the research on sentence matching can be applied to a large number of known natural language processing tasks,such as information retrieval,automatic question and answer,machine translation,dialogue system,paraphrase identification etc.The natural language processing tasks can be abstract into sentence matching problem to some extent,such as information retrieval can be summed up in the query and document matching,automatic answer can be summed up in questions and candidate answers matching,machine translation can be summed up in matching between two languages,the dialog system can be seen as a match between the previous conversation and the response,the paraphrase identification can be regarded as the match of two synonymous sentences.In a series of natural language processing tasks,we need to rely on the participation and collaboration of the sentence matching model.The performance of the sentence matching model can greatly affect the final performance of these natural language processing tasks.This paper studies the sentence matching model based on multi-perspective feature fusion,mainly doing the following three parts:(1)Improve the matching model from the perspective of word interaction,propose an attention mechanism based on the distribution of aspect-level sentiment difference,and combine the multi-round decision mechanism to perform multi-round word interaction matching.First of all,this paper proposes an attention mechanism based on the distribution of aspect-level sentiment difference to improve the interaction between cross-sentence words,and uses the sentiment space position perception vector to improve the interaction between intrasentence words,so that the model has the ability to perceive the subjective sentiment difference in the process of intra-sentence word interaction and cross-sentence word interaction.Then,this paper introduces a multi-round decision mechanism based on the accumulation of memory state,which iteratively updates the working memory state to make matching decisions in multiple rounds,so that the model can better understand the semantic of complex sentence.(2)Improve the matching model from the perspective of sentence encoding,combine syntactic graph convolution and self-attention mechanism to construct the sentence encoding.This paper uses the self-attention mechanism to extract the semantic information of sentence,and obtains high-quality sentence encoding through multi-dimensional feature extraction.In addition,in order to solve the problem that the self-attention mechanism is weak in the encoding of word position information,this paper combines the syntactic graph convolution to improve the position encoding mechanism and obtains structured position information in the process of sentence encoding.(3)The idea of word interaction and sentence encoding is combined to perform sentence matching by fusing the features of multiple perspectives.The disadvantage of the matching model based on word interaction is that it is weak in capturing the global feature and may lose some information about the overall feature of the sentence.The disadvantage of the matching model based on sentence encoding is that the semantic perception of local words is weak and the correspondence between some important words is easily ignored.In view of these shortcomings,this paper combines the advantages of the above two types of matching models,combines the methods of word interaction and sentence encoding,fuses the fine-grained features of words and the coarse-grained features of sentence,and joints multiple perspectives to perform sentence matching.The model can not only perceive the interaction between words and capture the local semantic features,but also understand the overall semantic of sentence from the global perspective,so as to obtain better matching result.