Extracting Methodological Sentences from Unstructured Abstracts of Academic Articles

Ruping Wang, Chengzhi Zhang, Yingyi Zhang, Jinzhu Zhang

Abstract: Methodological sentence is the smallest unit that depicts how the research method is used in one paper. Researchers can understand a method by reading these sentences. So, extracting methodological sentences automatically is meaningful for them to evaluate and select appropriate methods in their research process. However, previous studies rely too much on manually annotated corpus, in which the quantity is limited. Furthermore, some studies do not perform well when generalized to testing sets. In this paper, we use structured abstracts as training data to alleviate the burden of manually annotation. The label for each sentence is determined by its corresponding title in the abstract. Moreover, in order to extract methodological sentences more precisely, a rule-based method is applied for pruning the prediction result. In experimental results, the P, R, and F1 value after pruning are 65.14%, 57.00% and 60.80% respectively, which are all higher than those are not pruned.