IxaMed: Applying Freeling and a Perceptron Sequential Tagger at the Shared Task on Analyzing Clinical Texts

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 361–365, Dublin, Ireland, August 23-24, 2014. IxaMed: Applying Freeling and a Perceptron Sequential Tagger at the Shared Task on Analyzing Clinical Texts Koldo Gojenola, Maite Oronoz, Alicia P´erez, Arantza Casillas IXA Taldea (UPV-EHU) maite.oronoz@ehu.es http://ixa.si.ehu.es Abstract This paper presents the results of the Ix- aMed team at the SemEval-2014 Shared Task 7 on Analyzing Clinical Texts. We have developed three different sys- tems based on: a) exact match, b) a general-purpose morphosyntactic analyzer enriched with the SNOMED CT termi- nology content, and c) a perceptron se- quential tagger based on a Global Linear Model. The three individual systems re- sult in similar f-score while they vary in their precision and recall. We have also tried direct combinations of the individual systems, obtaining considerable improve- ments in performance. 1 Introduction This paper presents the results of the IxaMed team. The task is focused on the identification (Task A) and normalization (Task B) of diseases and disor- ders in clinical reports. We have developed three different systems based on: a) exact match, b) a general- purpose morphosyntactic analyzer enriched with the SNOMED CT terminology content, and c) a perceptron sequential tagger based on a Global Linear Model. The first system can be seen as a baseline that can be compared with other ap- proaches, while the other two represent two alter- native approaches based on knowledge organized in dictionaries/ontologies and machine learning, respectively. We also tried direct combinations of the individual systems, obtaining considerable im- provements in performance. These approaches are representative of different solutions that have been proposed in the literature This work is licensed under a Creative Commons At- tribution 4.0 International Licence. Page numbers and pro- ceedings footer are added by the organizers. License details: http://creativecommons.org/licenses/by/4.0/ (Pradhan et al., 2013), which can be broadly clas- sified in the following types: • Knowledge-based. This approach makes use of large-scale dictionaries and ontologies, that are sometimes integrated in general tools adapted to the clinical domain, as MetaMap (Aronson and Lang, 2010) and cTAKES (Xia et al., 2013). • Rule-based. For example, in (Wang and Akella, 2013) the authors show the use of a rule-based approach on the output of MetaMap. • Statistical techniques. These systems take a training set as input and apply different vari- ants of machine learning, such as sequen- tial taggers based on hidden Markov mod- els (HMMs) or conditional random fields (CRFs) (Zuccon et al., 2013; Bodnari et al., 2013; Gung, 2013; Hervas et al., 2013; Lea- man et al., 2013). • Combinations. These approaches try to take the advantages of different system types, us- ing methods such as voting or metaclassi- fiers (Liu et al., 2013). In the rest of the paper, we will first introduce the different systems that we have developed in section 2, presenting the main results in section 3, and ending with the main conclusions. 2 System Description The task of detecting diseases and their corre- sponding concept unique identifiers …

Content truncated to 500 words