Identifying Sections in Scientific Abstracts using Conditional Random Fields

Identifying Sections in Scientific Abstracts using Conditional Random Fields Kenji Hirohata† hirohata@nii.ac.jp Naoaki Okazaki† okazaki@is.s.u-tokyo.ac.jp Sophia Ananiadou‡ sophia.ananiadou@manchester.ac.uk †Graduate School of Information Science and Technology, University of Tokyo 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-8656, Japan Mitsuru Ishizuka† ishizuka@i.u-tokyo.ac.jp ‡School of Computer Science, University of Manchester National Centre for Text Mining (NaCTeM) Manchester Interdisciplinary Biocentre, 131 Princess Street, Manchester M1 7DN, UK Abstract OBJECTIVE: The prior knowledge about the rhetorical structure of scientific abstracts is useful for various text-mining tasks such as information extraction, information re- trieval, and automatic summarization. This paper presents a novel approach to cate- gorize sentences in scientific abstracts into four sections, objective, methods, results, and conclusions. METHOD: Formalizing the categorization task as a sequential label- ing problem, we employ Conditional Ran- dom Fields (CRFs) to annotate section la- bels into abstract sentences. The train- ing corpus is acquired automatically from Medline abstracts. RESULTS: The pro- posed method outperformed the previous approaches, achieving 95.5% per-sentence accuracy and 68.8% per-abstract accuracy. CONCLUSION: The experimental results showed that CRFs could model the rhetor- ical structure of abstracts more suitably. 1 Introduction Scientific abstracts are prone to share a similar rhetorical structure. For example, an abstract usu- ally begins with the description of background in- formation, and is followed by the target problem, solution to the problem, evaluation of the solution, and conclusion of the paper. Previous studies ob- served the typical move of rhetorical roles in sci- entific abstracts: problem, solution, evaluation, and conclusion (Graetz, 1985; Salanger-Meyer, 1990; Swales, 1990; Or˘asan, 2001). The American Na- tional Standard Institute (ANSI) recommends au- thors and editors of abstracts to state the purpose, methods, results, and conclusions presented in the documents (ANSI, 1979). The prior knowledge about the rhetorical structure of abstracts is useful to improve the performance of various text-mining tasks. Marcu (1999) proposed an extraction method for summarization that cap- tured the flow of text, based on Rhetorical Struc- ture Theory (RST). Some extraction methods make use of cue phrases (e.g., “in conclusion”, “our in- vestigation has shown that ...”), which suggest that the rhetorical role of sentences is to identify im- portant sentences (Edmundson, 1969; Paice, 1981). We can survey the problems, purposes, motivations, and previous approaches of a research field by read- ing texts in background sections of scientific papers. Tbahriti (2006) improved the performance of their information retrieval engine, giving more weight to sentences referring to purpose and conclusion. In this paper, we present a supervised machine- learning approach that categorizes sentences in sci- entific abstracts into four sections, objective, meth- ods, results, and conclusions. Figure 1 illustrates the task of this study. Given an unstructured ab- stract without section labels indicated by boldface type, the proposed method annotates section labels of each sentence. Assuming that this task is well formalized as a sequential labeling problem, we use Conditional Random Fields (CRFs) (Lafferty et al., 2001) to identify rhetorical roles in scientific ab- stracts.The proposed method outperforms previous approaches to this problem, achieving 95.5% per- 381 OBJECTIVE: This …

Content truncated to 500 words