Towards Guided Summarization of Scientific Articles: Selection of Important Update Sentences

Guided summarization is aimed to distinguish the information from the prior and later documents for a given domain that are read by readers. Update summary is a second-crucial component to be generated from this task, after the initial summary. The existing techniques of guided summarization for news article cannot be directly adapted to the scientific article due to the different location structure of important information between these domains. In this research, we try to conduct the selection of important update sentences which is part of steps in guided summarization for the domain of scientific articles. We employ and compare some selection algorithms, such as Maximum Marginal Relevance (MMR) and TextRank. Because there is an initial summary that will become a reference to generate non- redundant information for update summary, so we modified their algorithms to minimize the similarity between the initial summary and candidates sentences for update summary. The result shows that the feature of `TextRank+ROUGE-SU4' fully gives the best ROUGE-2 recall performance that is 15.21% and the modified MMR and TextRank equation is better than baseline.