Automatic sentence classifier using sentence ordering features for Event Based Medicine: Shared task system description

Automatic sentence classifier using sentence ordering features for Event Based Medicine: Shared task system description Spandana Gella University of Melbourne sgella@student.unimelb.edu.au Duong Thanh Long University of Melbourne lduong@student.unimelb.edu.au Abstract In this paper, we propose an automatic sen- tence classification model that can map sen- tences of a given biomedical abstract into set of pre-defined categories which are used for Evidence Based Medicine (EBM). In our model we explored the use of vari- ous lexical, structural and sequential fea- tures and worked with Conditional Random Fields (CRF) for classification. Results ob- tained with our proposed method show im- provement with respect to current state-of- the-art systems. We have participated in the ALTA shared task 2012 and our best per- forming model is ranked among top 5 sys- tems. 1 Introduction Evidence Based Medicine (EBM) or Evidence based practice is “systematically locating, ap- praising, and using contemporaneous research findings as the basis for clinical decisions” (Rosenberg and Donald, 1995). Considering the huge amounts of literature and millions of clin- ical articles currently available and continuously being added to databases like PubMed1, automat- ing the information access or searching scientific evidence for EBM is a crucial task. Currently evidence based practitioners use the PICO crite- rion which was proposed by Armstrong (1999) to construct queries and search information in EBM tasks. The PICO concepts or tag-sets are: Popu- lation (P), Intervention (I), Comparison (C) and Outcome (O). In this paper, we present a method that classi- fies sentences in the abstract of a clinical article 1http://en.wikipedia.org/wiki/PubMed according to PIBOSO criteria which is an exten- sion of PICO. PIBOSO has six tags: Population (P), Intervention (I), Background (B), Outcome (O), Study Design (SD) and Other (Oth). This in- formation could be used in constructing queries or searching relevant articles in the EBM task. A clear description of the PIBOSO tag-set is avail- able in (Kim et al., 2011), who proposed the tag- set. Our system is based on the CRF algorithm which was earlier used by Kim et al. (2011) for a similar task and proven to be useful. The major contribution of this paper is that we use a simple and large set of features such as lex- ical, structural and sequential features and show major improvements on the task of sentence clas- sification over earlier attempts. Our classification techniques have shown clear improvement over existing state-of-the art systems especially for un- structured abstracts. The paper is organised as follows: We present our related work in Section 2, describe the dataset for training and evaluation in Section 3, and our method and experimental setup in Section 4. We present the analysis of our results in Section 5 and conclude in Section 6. 2 Related work The first attempt to classify abstract sentences based on the PIBOSO schema is made by Kim et al. (2011). They used the Conditional Random Field (CRF) classifier for learning, and their fea- ture set included lexical features (unigram and bi- gram with part-of-speech), semantic features (us- ing metathesaurus), …

Content truncated to 500 words