UTU: Adapting Biomedical Event Extraction System to Disorder Attribute Detection

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 375–379, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics UTU: Adapting Biomedical Event Extraction System to Disorder Attribute Detection Kai Hakala University of Turku Graduate School (UTUGS), University of Turku, Finland Dept. of Information Technology, University of Turku, Finland kahaka@utu.fi Abstract In this paper we describe our entry to the Se- mEval 2015 clinical text analysis task. We participated only in the disorder attribute de- tection task 2a. Our main goal was to as- sess how well an information extraction sys- tem originally developed for a different task and domain can be utilized in this task. Our system, based on SVM and CRF classifiers, showed promising results, placing 3rd out of 6 participants in this task with performance of 0.857 measured in weighted accuracy, the of- ficial evaluation metric. 1 Introduction SemEval 2015 introduced a new subtask for the clin- ical text analysis track focusing on disorder mention attribute detection. These attributes describe the rel- evant information extracted from the textual context of the given disease mention, such as the severity or body location of the disease. The attributes were grouped into 9 separate categories, each with a pre- defined set of valid attribute classes. The task was defined as a template filling task where the textual cue words for the attributes have to be first identi- fied and then normalized to the correct class. Similar task with slightly different definition has previously been organized as part of the ShARe/CLEF eHealth shared task (Mowery et al., 2014). Due to time limitations we participated only in the task 2a in which the gold standard disorder mentions were given and only the attribute values had to be predicted. Our main motivation for this years en- try was to evaluate the performance of an existing information extraction system, TEES (Bj¨orne and Salakoski, 2013), previously developed for a dif- ferent domain and to assess how easily it can be adapted to a new task. 2 System Description Turku Event Extraction System (TEES) was origi- nally developed in 2009 for the BioNLP Shared Task on Event Extraction (Kim et al., 2009). This task fo- cused on the extraction of biological processes and interactions between genes and proteins (GGPs) de- scribed in biomedical literature. In this task each event, i.e. biological process or interaction, is rep- resented by a trigger word, which also describes the type of the event, and a set of argument GGP men- tions. The argument GGPs may also act in various roles, i.e. each argument is also typed. The par- ticipants were thus required to detect these trigger words, their types from a predefined set and the ar- guments, i.e. the relations between the trigger words and GGPs. Gold standard gene and protein mentions were provided by the organizers and consequently TEES does not include tools for named entity recog- nition, but presumes these to be given as input data. An example sentence along with the extracted event …

Content truncated to 500 words