Identifying Finding Sentences in Conclusion Subsections of Biomedical Abstracts

Segmenting scientific abstracts and full-text based on their rhetorical function is an essential task in text classification. Small rhetorical segments can be useful for fine-grained literature search, summarization, and comparison. Current effort has been focusing on segmenting documents into general sections such as introduction, method, and conclusion, and much less on the roles of individual sentences within the segments. For example, not all sentences in the conclusion section are describing research findings. In this work, we developed rule-based and machine learning methods and compared their performance in identifying the finding sentences in conclusion subsections of biomedical abstracts. 1100 conclusion subsections with observational and randomized clinical trials study designs covering five common health topics were sampled from PubMed to develop and evaluate the methods. The rule-based method and the bag-of-words based machine learning method both achieved high accuracy. The better performance by the simple rule-based approach shows that although advanced machine learning approaches could capture the main patterns, human expert may still outperform on such a specialized task.