Research On Automatic Text Summarization Based On Entity Information Embedding

With the rapid development of internet technology,a huge amount of data is generated on the internet every day.How to quickly and accurately extract the information that users need from these data has become a key task in the field of natural language processing.Automatic text summarization technology,as an important research direction of natural language processing,can automatically extract the main information from the original text,and has strong research value and application value.Currently,the field of automatic text summarization is mainly focused on the abstractive text summarization approach,which can generate summaries similar to those of human experts based on an understanding of the original text.However,the summaries generated by this approach are prone to inconsistencies with the original text.Related research has shown that such inconsistencies mainly occur in entity-related problems.To address this problem,this thesis proposes an abstractive text summarization model based on entity information embedding to solve the consistency issues that arise in text summarization.The main contents of this thesis include the following sections:(1)To address the problem of inconsistencies in entity-related facts due to insufficient use of entity information in abstractive text summarization,a summarization model based on entity-aware text encoding is proposed.By mixing entity and text encoding during the encoding process,entity information can be incorporated into the text encoding,allowing the entity information in the encoding to be used in the decoding process,reducing inconsistencies in entity-related facts in the model’s summaries.Finally,the effectiveness of the proposed model in improving summary fact consistency is verified through experiments.(2)To address the problem of entity-attribute mismatches in abstractive text summarization,a summarization model combining a fine-grained entity-attribute graph is proposed.By parsing the input text,entities and attributes are identified and connected into a fine-grained entity-attribute graph.The structure of the graph is conducive to expressing the ownership and connection relationships between entities and attributes.The thesis then uses a graph encoder to encode the entity-attribute graph and utilizes it in the encoding and decoding process of text summarization.By using the entity-attribute graph to assist summarization,the fact consistency of the summaries is improved by 3.34%according to the Fact CC score.(3)An automatic text summarization system is designed and implemented for news text summarization tasks.The system integrates the fact-consistent text summarization methods proposed in this thesis and can receive user input news text for content summarization,demonstrating the application value of abstractive text summarization technology.