Research On Text Summary And Evaluation Method Based On Neural Network

Automatic text summarization and summary evaluation are hot research directions in the field of natural language processing.The commonality of these studies is the model's understanding of the semantics of the original text.With the rapid development of science and technology,especially the development of machine learning and deep learning,these studies can be well applied to the user needs analysis of mechanical product design.There is a large amount of information in the customer's product requirements.The application of these technologies can accurately summarize the product design requirements and judge whether the automatic text summary model is accurate in analyzing the customer's requirements through evaluation methods.Abstract research mainly includes abstract generation and abstract evaluation.Among them,extractive and abstract abstracts are the two main methods of automatic text abstract generation.Extractive abstracts directly extract important sentences from the original text to form a summary,but with the development of neural networks,abstract abstracts have become the mainstream method of abstract generation.This method uses deep learning theory to train the model to understand the original text,and then generates an abstract.After the model generates the summary,it is necessary to evaluate the quality of the summary.The summary evaluation mainly includes external evaluation and internal evaluation.The internal evaluation method uses the information overlap between the reference abstract and the system abstract to evaluate the quality of the abstract.The external evaluation method is the opposite of the internal evaluation method.This evaluation method does not need to refer to the abstract.It uses the system abstract instead of the original document to execute a certain document application.Although both text abstracts and abstract evaluations have achieved good development,the abstract model accurately locates key semantic information and the mainstream internal abstract evaluation standard ROUGE in calculating deep semantic information between sentences to be evaluated and the evaluation results of this method There is still room for improvement without being affected by the subjectivity of the reference abstract,which can further promote the development of abstract and evaluation research.This paper studies the text summary and its evaluation method based on neural network.First,in order to solve the problem that the abstract model is difficult to locate and obtain key semantic information,which leads to inaccurate abstract generation,a summary model based on positioning attention and competition mechanism is proposed.This model improves the accuracy of abstract generation.Secondly,in order to solve the problem that the abstract evaluation method ROUGE only calculates the co-occurrence information between the system abstract and the reference abstract and does not consider the deep semantic information between the sentences to be evaluated and the evaluation results are subject to the subjectivity of the reference abstract,the abstract evaluation method is used The research is divided into two parts.First,a semantic matching model using selection gates and intra-class metric is proposed,which can solve the problem of low matching accuracy due to non-similar redundant features in abstract and original semantic matching.Second,a summary evaluation model using semantic similarity is proposed,and a semantic matching model using selection gates and intra-class metrics is combined with a feature fusion algorithm to improve the summary evaluation criteria.Aiming at the research of neural network-based text abstracts and their evaluation methodology,the prime contributions of this article are:First of all,to solve the problem that the abstract model is difficult to locate and obtain key semantic information,which leads to inaccurate abstract generation,this paper designs a summary model based on positioning attention and competition mechanism.The model mainly includes three parts: sequence-to-sequence network based on localization attention and flashback structure,selection gate encoder network,and competition mechanism.Firstly,the key words are used to assist key sentences by means of the probability superposition of important information,the attention mechanism is assigned to assign the weight of key information to locate the key semantic information,and it is input to the decoder to generate a summary.At the same time,the key sentence is input into the network based on the selection gate encoder,the key information is selected by calculating the semantic distribution probability of the words in the sentence,and the information is input into the decoder to generate a summary.Furthermore,a competition mechanism for calculating the abstract and the cosine value of the key sentence is designed,and the abstracts generated by the above two networks are optimized as the final result.Experiments show that the model is tested on LCSTS Chinese dataset with Rouge-1 of 38.17%,Rouge-2 of 22.24%,and Rouge-L of 34.97%,which is better than the current better RNN-distract and DRGD models.Secondly,in order to solve the problem of low matching accuracy caused by non-similar redundant features in abstract and original matching,this paper designs a semantic matching model using selection gates and intra-class metrics.First,the sentences to be compared are vectored into a selection gate after bidirectional long-term and short-term memory networks are vectored,and the key features closest to the original semantics are selected by calculating the semantic distribution probability of the words in the sentences to reduce the interference of redundant information.Secondly,a fusion loss method for learning fine-grained similar features is proposed.The distance between the semantic layer of the aggregation layer and the prediction layer is calculated,and it is fed back to the selection gate together with the original loss function.The inner distance is smaller,which is conducive to sentence semantic matching.Experiments show that the method is tested on Chinese-made data sets and Quora English data sets,and the accuracy rates can reach 96.01% and 88.67%,respectively,which is better than the current two-way multi-angle matching model with better performance.Finally,in order to solve the problem that the ROUGE abstract evaluation method only calculates the overlapping information between the system abstract and the reference abstract,and the evaluation result is affected by the subjectiveness of the reference abstract,which leads to the unreasonable abstract evaluation,this paper designs a summary evaluation model using semantic similarity.First,the abstract and the original text are input into the above semantic matching model using selection gates and intra-class metrics to obtain similarity scores.Secondly,this paper proposes an algorithm that integrates score features.The similarity score and ROUGE score are used to calculate the confidence distance matrix and relationship.Matrix,calculate the optimal fusion group by the limit value,and then use the maximum likelihood method to calculate the optimal fusion evaluation score of the optimal fusion group.Experiments show that this method is tested on the Chinese data set.The coefficients of variation F-1-cv are 55.593,58.488 and 71.462,and the coefficients of variation F-L-cv are 52.736,61.685 and 74.857.Tested on the English data set,the coefficients of variation F-1-cv were 105.621,42.626 and 43.124,and the coefficients of variation F-L-cv were 107.996,42.062 and 37.298,which also proved to be superior to the ROUGE evaluation method in extended experiments.In general,this article addresses the problems of inaccurate abstract generation and unreasonable abstract evaluation from several aspects,such as accurately positioning the key semantic information of the original text and improving the rationality of the ROUGE abstract evaluation method,and has achieved significant results for the future.The text summary technology is applied to the demand analysis of mechanical product design to summarize the design elements and provide technical support.