Research On Semantic Similarity Of Chinese Based On Improved Siamese Neural Network

Semantic similarity calculation has always been a hot and difficult topic in the research of natural language processing,and it is used as the main algorithm support in many computer applications such as search engine,intelligent customer service,translation software and so on.The emergence of deep learning has gradually replaced the traditional semantic similarity calculation.The Siamese neural network framework in deep learning has better performance in semantic similarity calculation due to its natural structural advantages.However,the feature extraction of sentence pairs is independent of each other,so the interactive features between two sentences cannot be obtained,and it is difficult to learn the logical relationship between sentences.In order to improve the above problems,this thesis proposes two improved models based on Siamese neural network to give full play to the advantages of Siamese neural network structure and improve the accuracy of semantic similarity calculation.(1)A semantic similarity interaction model based on Siamese ELECTRA network is proposed.Firstly,the coding layer of the Siamese neural network is constructed by using ELECTRA model,which can not only improve the computational efficiency of the model but also enable the model to obtain deeper semantic information and distance information between two texts.In order to extract the interaction features of two texts,the model combined BERT's next prediction task to build an interaction module,and used the obtained interaction features and text distance features to jointly measure semantic similarity.(2)The second improved model is semantic similarity model based on Siamese Uni LM network.The Uni LM network layer has a variety of encoding methods,which can be used as the coding layer of the Siamese neural network framework to fully obtain the semantic information of sentences and the context relationship between sentence pairs.Based on the two improved models proposed in this paper,experiments are carried out on the two text data sets of equivalence relation and question and answer relation respectively.Experimental results show that the accuracy of semantic similarity interaction model based on Siamese ELECTRA network is improved compared with the current advanced BERT model in both data sets.Moreover,it performs better on the equivalence task than the question and answer task.The semantic similarity model based on the Siamese Uni LM network performs as well as the BERT model in the equivalence relationship task.The accuracy of the semantic similarity model based on the Siamese ELECTRA network is 3.31% higher than that of the BERT model in the question and answer task,and 2.58% higher than that of the semantic similarity interaction model based on the Siamese ELECTRA network.