Research On Query Focused Summarization Generation Method

Text summarization generation technology can generate short and representative summary information from a given long text.Users can quickly obtain the information described in the document by reading these summaries.However,a document often contains multiple aspects of information.If you want to quickly obtain the aspect summary information you care about according to your interests,you need the support of query focused summarization technology.Query focused summarization generation is a technology for generating query-related summaries given documents and queries.In view of the characteristics of the query focused summarization generation task,this paper starts with the research on how to fully model the relationship between query and document and how to solve the query focused summarization generation task in the form of long document,proposes two query focused summarization generation methods,and designs a query focused summarization generation system based on these two methods.The main research work of this paper is as follows:(1)Query focused summarization generation based on crossing multi-head attentionIn the task of query focused summarization generation,deep level interaction modeling between queries and documents is required.To address this issue,a query focused summarization generation model is proposed,which uses the cross multi-head attention mechanism to fully model the interaction between document and query.At the encoding stage of the model,the embedded representation of documents,queries and summaries is initially constructed through the BERT pre-training model,and then the deep interactive semantic representation information between query and document is established through the cross multi-head attention mechanism between the multi-layer of Transformer Encoders.In the decoding stage of the model,the improved Transformer Decoder and multi-source Pointer-Generator network are used to generate the final summary,so that the generated summary has high expression coherence and semantic consistency with the document and query content.Finally,on two query focused summarization datasets,a comparative experiment with the existing summarization methods are carried out,and the experimental results verify the effectiveness of the model.(2)Query focused long document summarization generation based on key sentence extractionAiming at the problem of modeling long documents in query focused long document summarization generation task,a long document summarization generation model of "extract first and then generate" is proposed.The model consists of two parts: key sentence extractor and summary generator.In the key sentence extractor,the sentence-level multi-head self-attention mechanism is used to obtain the global dependency information of the long document,and then the output of the attention mechanism is mapped to a linear layer to obtain the key score of each sentence in the document.According to the key score,Top-k key sentences are extracted from the document.In the summary generator,the query and extracted key sentences are used as the input of the generative pre-training model to generate the final summary.Finally,the effectiveness of this model in processing query focused long document summarization generation is verified by comparative experiments.(3)Design and implementation of query focused summarization generation systemUsing above two query focused summarization generation methods,an online query focused summarization generation system is designed and implemented.The system adopts a front-end and back-end separation architecture,and provides two use methods of single data call and batch data call according to the needs of different types of users,which can help ordinary users and scientific researchers more easily use the models proposed in this paper.