CENTAL at SemEval-2016 Task 12: a linguistically fed CRF model for medical and temporal information extraction

Proceedings of SemEval-2016, pages 1286–1291, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics CENTAL at SemEval-2016 Task 12: A linguistically fed CRF model for medical and temporal information extraction Charlotte Hansart Damien De Meyere Patrick Watrin Andr´e Bittar C´edrick Fairon CENTAL Universit´e catholique de Louvain Place Blaise Pascal 1, 1348 Louvain-la-Neuve, Belgium Abstract In this paper, we describe the system de- veloped for our participation in the Clinical TempEval task of SemEval 2016 (task 12). Our team focused on the subtasks of span and attribute identification from raw text and proposed a system that integrates both statis- tical and linguistic approaches. Our system is based on Conditional Random Fields with high-precision linguistic features. 1 Introduction Extracting and linking temporal and medical infor- mation from medical documents is a highly useful task for many clinical applications and plays an im- portant role in health care assessment and patient safety (Sun et al., 2013). Since 2007, the SemEval competition includes a temporal information extrac- tion task which is now transposed onto the medical domain. This year, the twelfth task included six subtasks that are described in (Bethard et al., 2016). Our team focused on the first four of these (Phase 1 submis- sion), i.e. identification of entity spans and attribute values from raw text1. We present an approach combining linguistic rules and machine learning methods. The tools pre- sented here were initially developed for French to extract information (temporal expressions, Named Entities, etc.) from newspaper corpora, a text genre which is quite different from medical documents. 1For the event attributes, only the “polarity” attribute is han- dled by our system. This challenge was thus a good opportunity for us to evaluate the scalability of our tools for both a lan- guage other than French and a new text genre. 2 Data We received three different sets of medical docu- ments: two for the development of our system, and one for testing purposes. The documents are clinical notes and pathology reports. The Train dataset com- prises 297 clinical reports. The Dev dataset, used for testing our developments, contains 150 reports. Finally, the Test dataset, available only for the eval- uation of our model, includes 153 documents. All the documents are manually de-identified by the Mayo Clinic and annotated according to an ex- tension of the ISO-TimeML standard (Pustejovsky et al., 2010): the THYME Annotation Guidelines, developed for the THYME project (Styler et al., 2014). 3 Methodology The originality of our approach comes from the simultaneous consideration of both terminological and linguistic resources which feed a statistical model based on the Conditional Random Fields paradigm (Lafferty et al., 2001). In our opinion, this strategy allowed us to take full advantage of the precision inherent to symbolic ap- proaches while enabling us to benefit from the flex- ibility of supervised statistical modeling methods. Such a hybrid methodology has proven useful in pre- vious research dealing with part-of-speech tagging (Constant and Sigogne, 2011). 1286 3.1 Statistical analysis Models such as Conditional Random Fields …

Content truncated to 500 words