Cross-lingual Text Summarization Model Research Based On Deep Learning

With the rapid development of human science and technology,artificial intelligence technology has been widely used in the current era of big data,accompanied by the explosive growth of massive data containing rich human social behavior information,including network public opinion,current affairs news,user comments and other text data,how to accurately and effectively extract and refine key information from these massive data according to the needs of users is an urgent difficulty to be solved.Text summarization technology was born in the context of solving the above problems.It mainly builds a model to understand and program the contextual semantic information in the text,while reducing redundant information,refining key information,or migrating to use NLP methods to generate summaries corresponding to the original text.Most of the current text summarization research focuses on the monolingual text summarization task dominated by English text,and many different text summarization models have been proposed and built around this task.Although the existing text summarization models have good performance and the quality of the generated summaries is also high,but it still faces the following shortcomings: First,most of the existing studies focus on the performance improvement of a specific algorithm or model for text summarization tasks,and few studies have experimentally tested and compared the effects of different types of deep learning models in text summarization work.Secondly,the research work on single-language text summarization is the mainstream of the current related academic fields,while the research resources devoted to cross-language text summarization work are relatively few.However,there are natural application defects in single-language text summarization.In a specific application,the language of the source document and the abstract is the same.In today’s increasingly globalized and informatized world,urgent problems such as understanding and transmission of cross-language information cannot be fundamentally solved.In view of the above two shortcomings,this thesis conducts research from the following two aspects:1.For question 1,few studies have experimentally examined the differences in the performance of different kinds of deep learning models in text summarization work.This thesis divides the current deep learning methods applied to text summarization into two categories(sequence-to-sequence models and pre-trained models),and compares their performance on specific public datasets.For the sequence-to-sequence model,this thesis adopts the Seq2 Seq and its variant architecture,which consists of an encoder and a decoder,and implements extractive text summarization and generative text summarization according to algorithm adjustment.For the pre-training model,this thesis adopts Transformer,BERT and its variant architecture.This thesis uses the English public dataset CNN/Daily mail to conduct experiments on these two types of algorithm models,and compare their performance and final summary quality in extractive summarization and generative summarization.The experimental results show that most generative summarization models have better performance and summary quality than extractive summarization models.Among the sequence-to-sequence algorithm models,the Conv Seq2 Seq model using CNN structure has the best performance and summary quality.In the pre-training algorithm model,the BERT+Transformer model with the encoder selected as BERT and the decoder selected as Transformer has the best performance and summary quality.2.For question 2,there are too few research resources devoted to cross-language text summarization work compared to monolingual ones.This thesis explores text summarization in cross-language application scenarios,and proposes a cross-language text summarization model CLTSTP(Cross-Lingual Text Summarization Based on Translation Pattern)based on translation transformation patterns.Specifically,the model proposed in this thesis divides the cross-language text summarization work into three stages: source word attention,crosslanguage word translation,and summary word generation.At the same time,three different translation transformation modes,"Original","Equal" and "Adjust",are adopted in the crosslanguage word translation stage.Among them,source word attention can help the model better combine the context and focus on key words in the source document,cross-language word translation can help the model better understand the semantic differences between different languages,and the summary word generation can prompt the model to select summary words with more complete semantic information.This thesis uses the ChineseEnglish public cross-language dataset ZH-EN and the English-Chinese public crosslanguage dataset EN-ZH to conduct experiments on the model constructed above.The experimental results show that the cross-language text summarization model based on translation conversion mode established in this thesis has better performance and summary quality than the selected baseline methods,and the model with "Adjust" translation conversion mode has the best results.Its ROUGE scores are 29.56%,14.12% and 26.42%,respectively.After the above exploration,this thesis experimentally tests and compares the performance and summary quality of two types of text summarization methods based on deep learning algorithms—sequence-to-sequence models and pre-trained models in singlelanguage text summarization.Further,this thesis is not limited to single-language application scenarios,but goes deep into cross-language application scenarios,and establishes a crosslanguage text summarization model with good performance and summary quality by integrating the translation transformation mode.It provides theoretical and practical reference for related research.