Research On Automatically Generated Text Summary Evaluation Based On Deep Learning

With the rapid development of the Internet and the explosive growth of information,text information is the main source for people to obtain information.In recent years,natural language processing technology has received more and more attention from researchers.The automatic generation of text summarization is an important technology Branch.This paper researches on the technology of automatic text summarization,and takes sentence semantic similarity matching technology and abstract evaluation technology as supporting research of abstract technology.First,by improving the traditional attention mechanism and loss function,a summary model based on the semantic similar attention mechanism is proposed.The semantic feature vector cosine similarity distance of the original text and the generated summary is fused with the conditional probability loss function to feed back to the attention.Mechanism to make the generated abstract closer to the original semantics.Secondly,by studying the application of semantic similarity in summary technology and summary evaluation,a sentence semantic similarity matching model based on optimal features is proposed to judge the semantic similarity of two sentences.Finally,on the basis of the semantic similarity matching model,by combining the Rouge summary evaluation algorithm,a summary evaluation method based on optimal fusion score is proposed,which makes the evaluation result closer to manual evaluation.The main contributions of this article are as follows:1.For the current summary model,it is difficult to obtain context information related to the semantics of the original text,resulting in the problem of inaccurate summary generation.This paper proposes a summary model based on semantic similar attention mechanism.First,a semantic similarity attention mechanism is established.By calculating the cosine distance between the encoder and decoder hidden layer states,and introducing the distance into the attention mechanism to assist the mechanism to obtain context semantics with high semantic similarity to the original text information.Secondly,a fusion loss method for learning semantic similarity features is proposed,and the weighted fusion of the cosine distance and the original loss function is fed back to the semantic similarity attention mechanism,and then the context information is updated iteratively.Finally,the decoder uses the semantic information closest to the original text to generate a more accurate summary.The experimental results show that the model has an increase of 0.65% and 0.17% in the evaluation scores of Rouge-1 and Rouge-2.2.Optimal features selection is essential for sentence similarity recognition.In this paper,a semantic matching model based on optimal selection mechanism is designed.Compared with other models that without using an optimal feature selection mechanism,the model uses optimal features that are closest to the original semantics to match,resulting in higher matching.This paper proposes an improved semantic information selection mechanism.This mechanism selects a candidate feature that is as close as possible to the semantics of the original text by calculating the probability of distribution of semantic information in the sentence.Furthermore,a new attention focus mechanism is designed,which iteratively updates the information weight to obtain another candidate feature that may affect the matching accuracy.Then,an optimal feature selection algorithm is designed by using the context vector and two candidate features.The coefficients are selected,and the optimal features are further selected from the candidate features for semantic matching.The experimental results show that the matching accuracy of the model on the English and Chinese data sets is significantly improved by 2.07% and 0.86%.This model is better than the current matching model with better performance.3.The Rouge evaluation method does not consider the semantic correlation between the original text and the generated abstract,and only calculates the score of the coincident unit between the two,resulting in the problem of unreasonable evaluation assessment method.First,this paper proposes a sentence semantic similarity matching model with a selection gate.The key features of both the original text and the generated text are obtained by calculating the probability distribution of words.This feature is used to match sentence similarity and obtain a semantic similarity score.Secondly,the Rouge-1,Rouge-2 and Rouge-L methods are used to calculate the scores of overlapping units between the original text and the generated text,respectively.Finally,by calculating the confidence distance matrix and correlation matrix of the semantic similarity score and Rouge score,the optimal fusion group of the two is obtained,and then the optimal fusion data is obtained using the principle of maximum value.Experiments show that the method in this paper uses Chinese data sets and English data sets to evaluate the abstracts generated by different models,which are close to manual evaluation standards.In summary,the research content of these three aspects is a complete system.The summary model based on the semantic similar attention mechanism is the core,and a summary evaluation method based on the optimal fusion score is verification.The sentence similarity matching model based on the optimal selection mechanism is an important part of both.