Challenges in Information Extraction from Tables in Biomedical Research Publications: a Dataset Analysis

Challenges in Information Extraction from Tables in Biomedical Research Publications: a Dataset Analysis Tatyana Shmanina1,2, Lawrence Cavedon3, Ingrid Zukerman1,2 1Clayton School of Information Technology, Monash University, Australia 2NICTA Victoria Research Laboratory, Melbourne, Australia 3School of Computer Science and IT, RMIT University, Australia 1firstname.lastname@monash.edu, 3firstname.lastname@rmit.edu.au Abstract We present a study of a dataset of ta- bles from biomedical research publica- tions. Our aim is to identify characteris- tics of biomedical tables that pose chal- lenges for the task of extracting informa- tion from tables, and to determine which parts of research papers typically contain information that is useful for this task. Our results indicate that biomedical tables are hard to interpret without their source pa- pers due to the brevity of the entries in the tables. In many cases, unstructured text segments, such as table titles, footnotes and non-table prose discussing a table, are required to interpret the table’s entries. 1 Introduction Automation of information extraction (IE) from biomedical literature has become an important task (Shatkay and Craven, 2012). In particu- lar, biomedical IE enables the semi-automation of tasks such as document indexing (Aronson et al., 2004) and database curation, e.g., (Donaldson et al., 2003; Karamanis et al., 2008). Most research in biomedical IE has concen- trated on information extraction from prose. How- ever, much important data, such as experimental results and relations between biomedical entities, often appear only in tables (Ansari et al., 2013). This insight was confirmed experimentally for the task of mutation database curation. In particu- lar, Wong et al. (2009) showed that for a sam- ple of research articles used to populate the Mis- match Repair database (Woods et al., 2007), tables served as a sole source of information about mu- tations for 59% of the documents. Yepes and Ver- spoor (2013) reported that a text mining tool ap- plied to full articles and their supplementary ma- terial, used to catalogue mutations in the COS- MIC (Bamford et al., 2004) and InSiGHT (Plazzer et al., 2013) databases, could recover only 3-8% of the mutations if only prose was considered. An additional 1% of the mutations was extracted from tables in the papers, with an improvement of mu- tation coverage to about 50% when supplementary material (mostly tables) was considered. Information extraction from tables (Table IE) comprises various tasks, such as (1) classification of table entries or columns into a set of specific classes (Quercini and Reynaud, 2013; Wong et al., 2009); (2) association of table entries or columns with concepts from a domain vocabulary (Assem et al., 2010; Yosef et al., 2011); and (3) extraction of relations, defined in a vocabulary, between enti- ties in tables – usually done with Task 2 (Hignette et al., 2009; Limaye et al., 2010; Mulwad et al., 2013; Venetis et al., 2011). These tasks are of- ten performed by consulting external knowledge sources. However, despite the intuition that un- structured text accompanying tables often pro- vides helpful information, little use has been made of such text. Examples of such usage are …

Content truncated to 500 words