Using Argumentative Zones for Extractive Summarization of Scientific Articles

摘要

Information structure, i.e the way speakers construct sentences to present new information in the context of old, can capture rich linguistic information about the discourse structure of scientific documents. Information structure has been found useful for important Natural Language Processing (NLP) tasks, such as information retrieval and extraction. Since scientific articles typically follow a certain discourse structure describing the prior work, problem being solved, methods used, and so forth, it could also be useful for summarization of these articles. In this work we focus on a scheme of information structure called Argumentative Zoning (AZ), and investigate whether its categories could support extractive text summarization in a scientific domain. We develop a summarization system that uses AZ categories (ⅰ) as features and (ⅱ) in the final sentence selection process. We evaluate the system directly as well as using task-based evaluation. The results show that AZ can support both full document and customized summarization. We report a statistically significant improvement in summarization performance against a competitive baseline that uses journal section labels instead of AZ information.