Research On Semantic Matching Of Single Round Dialogue Algorithm Based On Deep Learning

In recent years,combat simulation technology has developed rapidly,and it has been used to assist in command training,which has greatly improved the organizational efficiency and training effect of personnel training.In particular,simulation training based on equipment such as VR/AR soldier wearables,training simulators,etc.is an important way to conduct personnel training.Voice interaction is the main way in the human-computer interaction mode,and it is also the basis for supporting simulation training.The voice interaction in simulation training is to realize the interaction between Computer-Generated Forces(CGF)and real forces through voice commands,in order to jointly accomplish cooperation and confrontation tasks.In the single-round dialogue scene that is common in the battlefield,CGF understands the voice command issued by the real forces and responds,which can effectively improve the fidelity of the training.The technical difficulty lies in identifying the semantic information,and judging whether the CGF's answer is correct by calculating the similarity between the two sentences,and the CGF completes the related tasks.The need to measure text similarity at the semantic level can be attributed to semantic matching tasks.Semantic matching needs to calculate the matching degree of sentence pairs on the semantic level,involving many difficulties such as word representation,sentence representation and sentence pair representation.Most of the traditional methods are based on lexical features,which are completely accomplished by manually extracting the linguistic features of specific tasks.There are many disadvantages such as difficulty in obtaining language tools,difficulty in extracting sentence features,less learning parameters,and poor generalization ability.The deep learning method based on neural network can automatically extract features from the original data,avoiding many shortcomings of traditional methods,and can effectively deal with semantic matching problems.Based on the method of deep learning,this paper proposes a multi-granularity convolution neural network.By focusing on the local information of sentences,the algorithm can obtain linguistic features of different granularity such as vocabulary,phrase and even sentences,so as to obtain a more comprehensive sentence representation.In addition,it is considered that semantic matching should fuse information between sentence pairs rather than modeling sentences separately.In this paper,a two-direction attention mechanism is proposed.Through the soft Alignment method,the semantic information is merged together,then the word weight with strong semantic relevance is enhanced,and the weight of words with smaller semantic relevance is reduced.And then the algorithm generates the sentence pair representations with semantic associations.Finally,this paper combines these two methods to provide a new solution for semantic matching.The experimental results on the three datasets show that the multi-granularity convolution neural network increases by 8% on average compared to the single-layer convolutional neural network on the two commonly used measures of MAP and MRR.The two-direction attention mechanism increased by an average of 7% relative to the pooled attention mechanism on the measures of MAP and MRR.The semantic matching algorithm of multi-granularity convolution neural network with two-direction attention mechanism formed by the combination of the two is improved by about 3% compared with the current classical semantic matching algorithm on the same measures.This fully proves the effectiveness of the proposed algorithm and lays a foundation for later engineering applications.