Research On Discrimination Method Of Semantic Textual Similarity Based On Neural Network

Text similarity discrimination is one of the basic technologies of natural language processing tasks,such as machine translation,automatic question answering.At present,the text similarity discrimination method still has problems to be solved in the following aspects: the sentence focuses and word semantics,which play an important role in the semantic expression of natural language,cannot be well integrated into the similarity discrimination,and the interactive features between texts are not fully utilized.These problems lead the model to have limited understanding of the true semantics contained in the text,and the results of similarity discrimination often cannot meet the needs of practical applications.In response to the above problems,based on the siamese network model,this paper integrates the idea of interactive matching,and proposes a multiple interactive matching model combined with natural language semantic expression characteristics to discriminate the semantic textual similarity.The work in this paper has been supported by the National Key R & D Program Project "Internal and External Connected Trial Execution and Litigation Services Collaborative Support Technology Research"(2018YFC0831300).The main work of this paper is as follows:(1)Aiming at the problem that the existing semantic textual similarity discrimination models ignore the characteristics of natural language semantic expression,this paper proposes a direct matching method based on sentence focuses recognition and word semantics interaction.With the help of the part-of-speech tagging tool and the ontology knowledge(How Net),this method completes the introduction of the information of sentence focuses and word semantics,and also enables the extraction of early interactive features at the word level when introducing semantic information to the model.Experimental results show that,compared with the basic model,the accuracy of the model built based on this method is improved by 4.29% for the semantic textual similarity task.(2)For the problem of insufficient interactive features extraction,in addition to the aforementioned method that has achieved early interactive features extraction of words,this paper also proposes an indirect interactive matching attention method based on siamese Bi LSTM.It is oriented to the context feature vector output by the neural network model,and further extracts the interactive features of the text pair.In this method,drawing on the idea of interactive matching,this paper designs a differential attention mechanism to introduce the differences between context feature vectors into the expression of sentence vectors in the form of attention.In this way,the local difference is increased to the global difference,thus the text representation will be more conducive to the semantic textual similarity task.Experimental results show that,compared with the basic model,the accuracy of the model built based on this method is improved by 1.55% for the semantic textual similarity task.Combining the above work with the basic model,the final model MIM-SEC has been built.In order to verify the effectiveness of the model,a control experiment and a contrast experiment have been conducted in this paper.The former proves that the existence of the main structure of the model is effective and reasonable and its accuracy on the dataset is improved by 4.81% compared to the basic model,the latter shows that the performance of MIM-SEC on the dataset is better than the common semantic textual similarity discrimination models.Therefore,it can be concluded that the multiple interactive matching model combined with natural language semantic expression characteristics proposed in this paper is an effective discrimination method of semantic textual similarity based on neural network.