Summarizing Scientific Articles By An Improved Discourse-Aware Attention Model

With the amount of information is getting larger and larger,the scientific articles have been published rapidly in especially,we would find it difficulties to get the overall content among these documents.Some researches have paid attention to this problem,however,two main problems exsist in these researches.Firstly,as we known,the scientific articles mainly contain the conponents such as the,introduction,all level titles,main body,conclusion and references.Some recent studies have explored the chapter discourse-aware method which takes some of mentioned conponents into consideration to extract the information from the scientific articles,but the conponents are partial consideration that such as all level titles,references have not been considered in their method.Furthermore,the discourse feature in paragraphes also have not been taken into account.Secondly,as the ordinary summarization evaluation methods such as the Edmundson and ROUGE still focus on the coverage degree in word level.But Scientific Article Sum-marization which would contain a lot of noise and knowledge in articles should be paid more attention,the ordinary abstract evaluation indexes are not suitable for evaluating the summarization of scientific articles to some extent.In order to handle these two problems in summarizing scientific articles above,two main methods would be stated in the following parts:On one hand,more comprehensive discourse-aware of scientific articles would be taken into account in this thesis.Summarization of the scientific articles would be performed in hierarchical methods by considering the improved chapter and paragraph discourse structure in detail.As for the improved chapter discourse structure,all level titles and references that contain more general information about the articles would be added into the chapter discourse-aware method.What'more,this study would put discourse-aware attention into the paragraphs.As we known,for the stable structure of scientific articles,the first and last sentence would cover most information than other sentences in paragraph,so more attention would be weighted on it.On the other hand,the summarization evaluation method which is based on the knowl-edge would be approved in this thesis that the knowledge coverage degree would instead the word coverage degree to improve the evaluation effect for summarizing scientific articles.Therefore,the major innovation points in this study would focus on the following three parts:a).All level titles and references that contain more general information about the scientific articles would be added into the chapter discourse-aware method.b).The paragraph discourse-aware method would be proposed to promote the effect of summarization.c).A evaluation method which is based on the knowledge coverage degree would be approved in this thesis.Some experiments have been performed in this researches.For measuring the effect of the improved discourse-aware models,two evaluation indexes such as the value of ROUGE and KBEI(Knowledge Based Evaluation Index)both haven been used.Comparing the previous researches with the ROUGE value,the results show that we have gotten a better result than some previous researches.What's more,the KBEI(Knowledge Based Evaluation Index)method reduces the noise in evaluating the summarization of scientific articles by comparing with the named entities between the reference summarization and machine generation summarization.