Multi-class SVM for Relation Extraction from Clinical Reports

Proceedings of Recent Advances in Natural Language Processing, pages 604–609, Hissar, Bulgaria, 12-14 September 2011. Multi-Class SVM for Relation Extraction from Clinical Reports Anne-Lyse Minard1, 2 Anne-Laure Ligozat1, 3 Brigitte Grau1, 3 (1) LIMSI-CNRS, BP 133, 91403 Orsay Cedex, France (2) Universit´e Paris-Sud, 91400 Orsay, France (3) ENSIIE, 1 square de la r´esistance, 91000 ´Evry, France firstname.lastname@limsi.fr Abstract Information extraction in specialized texts raises different problems related to the kind of searched information. In this pa- per, we are interested in relation identifi- cation between some concepts in medical reports, task that was evaluated in the i2b2 2010 challenge. As relations are expressed in natural language with a great variety of forms, we proceeded to sentence analysis by extracting features that enable all to- gether to identify a relation and we mod- eled this task as a multi-class classification based on an SVM, each type of relation representing a class. We will present the selection of the features used by our sys- tem and an error analysis. This approach allowed us to obtain an F-measure of 0.70, classifying the system among the best sys- tems. 1 Introduction Medical information systems have developed past years, and are used for the storage of the infor- mation to facilitate the access to data, to help to search medical information about the patient or to provide decision support to improve the quality of care. The information processed mainly concern medical literature and medical records of patients, such as the clinical reports and the consultation re- ports which contain a lot of information about the medical follow-up. A large part of this informa- tion is in texts. So an important issue is to convert automatically all this information into some struc- tured knowledge as it is the starting point for the development of some semantic interrogation tools and high level processing of this information. Extraction of medical information raises dif- ferent problems related to the kind of informa- tion sought in texts: i) the recognition of med- ical terms, ii) related concepts and iii) relations between them. A terminologic analysis of doc- uments lead to build semantic indexes used to search information (Jonquet et al., 2010). Iden- tifying relations between concepts provides a more structured representation. That is useful for precise information retrieval, for example for Question-Answering systems (Tjongkimsang et al., 2005), (Embarek and Ferret, 2010). In this paper, we present our work1 on the iden- tification of relations in clinical reports, task of the i2b2 2010 challenge2. One of the goals of the challenge was to identify several kinds of relations between concepts (treatment, test and problem). These relations are expressed in the reports by a wide range of wordings. The incompleteness of semantic knowledge bases combined with the dif- ficulty of relating wordings with conceptual rep- resentations is an obstacle to the realization of a deep analysis of sentences which would highlight the relations between concepts. Thus, we considered that a lot of sentence char- acteristics such as the words used, their syntactic …

Content truncated to 500 words