Analysis And Research On Factual Abstractive Automatic Summarization And Factness Metrics

Automatic summarization is an important task in natural language processing,and its goal is to compress text,convert it into sentences with key information,and use it as summary.With the in-depth study of automatic summarization,some researchers also focus on the factual research in summarization,which focuses on the set of key semantic information in the text.In recent research on summary factness,many researchers add factrelated attention to the decoder of the model.But this approach does not allow the model to truly understand the factnesss which makes it fail to make full use of the information contained in facts.At the same time,because of the research of facts pays more attention to the semantic connection between summary and the original text,the ROUGE metric alone cannot fully evaluate the quality of the facts in the summary.Therefore,reasonable metric related to factness also need to be proposed urgently.To sum up,this thesis explores the factuality in abstractive automatic summarization from two perspectives,aiming at the imperfection of current summary factual research,namely:From the perspective of model,a fact-based abstractive summary model is designed,by encoding the facts at the encoder,the model can understand the facts from the semantic level,and at the same time,by adding an attention mechanism on the decoder,the facts are emphasized again.Experiments have verified that after using the Bart pre-training model,the summary generated has surpassed most summary models in the score of the ROUGE metric.From the perspective of metrics,this thesis designs new factual metrics,and evaluates the degree of agreement between the summary and the factual information in the original text from the perspective of semantics and structure,this thesis also designs a series of comparative experiments to verify that the new metrics are closer to the manual evaluation,and are better than the traditional summary metrics.Experiments show that incorporating facts into the encoder can effectively improve the quality of the summary and improve the consistency between the facts in the summary and the original text.The model scores as high as 19.11%and 37.45%on ROUGE-2 and ROUGE-L metrics,which indicates that the factual information that this model focuses on has a positive impact on the generation of summary.At the same time,it is found that the metric obtained by reasonably combining the metrics of semantics and syntactic structure are more complete,and compared with ROUGE,the factuality can be evaluated more comprehensively.