Research On Data To Text Generation Based On Deep Learning

Data-to-text generation is an important research direction in the field of natural language processing,which aims to generate logically clear and semantically fluent text from structured data.With the continuous acceleration of the social informatization process,the field of deep learning is becoming more and more popular,and the society has gradually moved from the traditional Internet field to the development of the concept of the metaverse.Therefore,in terms of text writing,people hope that machines can automatically write high-level text like humans,which greatly improves the efficiency while ensuring the quality of the text.However,there is currently a lack of such research on data-to-text generation,and the landing results are even rarer.Therefore,research in this field has great theoretical and practical application value.Although the data-to-text generation task has achieved certain results in recent years,there are still problems to be solved.For example,it is difficult to select key content for description in the huge structured data.Meanwhile the fluency and coherence of the generated text are poor,there are even problems with repetitive text.Therefore,In this paper,we conduct research through deep learning methods to try to solve the above problem.First of all,this paper conducts research on semantic matching.Semantics is the soul of natural language.In order to do a good job in generating tasks,the judgment of semantics needs to be correct.In order to solve the problem of sufficient semantic information interaction between sentence pairs,a semantic matching method based on BERT and dense composite network is proposed.On the basis of studying semantics,a data-to-text generation method based on selective coding and fusion of semantic loss is proposed,which solves the problems of difficulty in selecting key content and repetitive texts in the process of text generation,and improves text coherence.Finally,after various data tests,it is proved that the method in this paper has the best performance.The contributions of this paper are as follows:1.A semantic matching method based on BERT and dense composite network is proposed.Through the dense connection between BERT embedding and composite network,the accuracy of text semantic matching is significantly improved.First,using BERT as a word embedding model,the high-quality word vector representation is obtained through iterative feedback during the training process,and then high-quality sentence pair semantic information is obtained.Secondly,a dense composite network is designed.Bi-LSTM first obtains the global semantic information of sentence pairs,and then Text CNN extracts and integrates local semantic information to obtain the key features of each sentence and the correspondence between sentence pairs,and the BERT Fusion with the hidden output of Bi-LSTM and the pooled output of Text CNN.Finally,summarizing the association state between networks during the training process can effectively prevent network degradation and enhance the model’s judgment ability.The experimental results show that on the community question answering(CQA)long text dataset,the method in this paper has a significant effect,with an average improvement of 45%.Both the method and performance are due to the strong baseline model,mainly in:(1)The BERT pre-training model is used to represent the word embedding,which not only encodes and represents the word itself,but also obtains the position information of the word and the paragraph information of the sentence,which makes the semantic expression of the sentence more complete and the level more clear.(2)Design a dense composite network composed of Bi-LSTM and Text CNN,integrate the sequence information of sentences and local key semantic information,and connect the output of the hidden layer and the output of the pooling layer at the same time,so that the model can summarize the learning state before and after,and enhance the judgment ability.2.A method with data to text generation based on selecting encoding and fusing semantic loss.By highlighting key content and reducing the redundancy of text description information,the quality of the generated text is significantly improved.First,a new selection network is designed,which uses the amount of information related to data records as the coding basis for content importance,and multiple rounds of dynamic iterations of the results to achieve accurate and comprehensive selection of important information.Secondly,in the decoding process using long-term short-term memory(LSTM),a hierarchical attention mechanism is designed to assign dynamic selection weights to different entities and their attributes in the hidden layer output to obtain the best generated text Recall rate.Finally,a method of calculating the semantic similarity loss between the generated text and the reference text is introduced.By calculating the cosine distance of the semantic vectors of the two and iteratively feedback to the training process to obtain the optimization of key features,while reducing the redundancy of description information and improving the model BLEU performance.The experimental results show that the test Precision rate,Recall rate and BLEU is up to 94.58%,53.72% and 17.24,which are better than existing models:(1)This method can effectively pay attention to important words first through the selection network designed by the encoding layer,and then use such words to participate in the decoding output during decoding.(2)On the basis of the original logarithmic loss,the semantic similarity loss is introduced.Through the joint optimization of the two losses,the model’s ability to control the global semantics is improved,thereby improving the fluency of the generated text and solving the text repetition to a great extent.