ULISBOA at SemEval-2016 Task 12: Extraction of temporal expressions, clinical events and relations using IBEnt

Proceedings of SemEval-2016, pages 1263–1267, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics ULISBOA at SemEval-2016 Task 12: Extraction of temporal expressions, clinical events and relations using IBEnt Marcia Barros, Andre Lamurias, Gonc¸alo Figueiro, Marta Antunes Joana Teixeira, Alexandre Pinheiro and Francisco M. Couto LaSIGE, Faculdade de Ciˆencias, Universidade de Lisboa, Portugal mbarros@lasige.di.fc.ul.pt Abstract This paper describes our approach to partici- pate on SemEval2016 Task12: Clinical Tem- pEval. Our system was based on IBEnt, a framework to identify chemical entities and their relations in text using machine learning techniques. This system has two modules, one to identify chemical entities, and other to identify the pairs of entities that represent a chemical interaction in the same text. In this work we adapted both IBEnt modules to ex- tract temporal expressions, event expressions and relations, by creating new CRF classifiers, lists and rules. The top result of our system was in phase2 for the identification of narra- tive container relations where it obtained the maximum score of precision (0.823) from all participants. 1 Introduction In this paper we present our approach to partici- pate on SemEval 2016 Task12: Clinical TempEval (Bethard et al., 2016) challenge. In phase1, par- ticipants had only access to the raw text, and they were asked to identify and classify time expres- sions (TIMEX3), event expressions (EVENT) and relations between expressions (RELATION), and in phase2 participants had access to the raw text and the manual EVENT and TIMEX3 annotations, and they were asked to identify only RELATION anno- tations. Our team participated in TS (identify the span of TIMEX3 expressions), ES (identify the span of EVENT expressions) and CR (identify narrative container RELATION) subtasks in phase1, and in phase2 we participated in the CR subtask. Our sys- tem was based on IBEnt (Lamurias et al., 2015), a framework that identifies chemical entities and their relations in text using machine learning techniques. Although IBEnt has been designed to extract chem- ical entities, we wanted to find out its potential to deal with other type of expressions. IBEnt has two modules: module one, an improvement of the tool developed by (Grego and Couto, 2013), recognizes chemical entities based on the Stanford NER soft- ware (Finkel et al., 2005) to train Conditional Ran- dom Field (CRF) classifiers using labelled data as input; module two identifies the pairs of entities that represent a chemical interaction in a given text based on machine learning techniques and domain knowl- edge. In particular, this latter module uses a non- linear kernel, the Shallow Language kernel (Giu- liano et al., 2006) taking into account both the global and local context of each entity to determine if they are interacting or not. The Shallow Language Ker- nel uses the word, lemma, POS and tag of each to- ken to train a classifier. One instance was generated for each candidate pair, whose elements were iden- tified with a specific tag. In our system we modified both modules in order to extract the relevant expres- sions …

Content truncated to 500 words