CDE-IIITH at SemEval-2016 Task 12: Extraction of Temporal Information from Clinical documents using Machine Learning techniques

Proceedings of SemEval-2016, pages 1237–1240, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics CDE-IIITH at SemEval-2016 Task 12: Extraction of Temporal Information from Clinical documents using Machine Learning techniques Veera Raghavendra Chikka Center For Data Engineering International Institute of Information Technology Hyderabad, India raghavendra.ch@research.iiit.ac.in Abstract In this paper, we demonstrate our approach for identification of events, time expressions and temporal relations among them. This work was carried out as part of SemEval-2016 Challenge Task 12: Clinical TempEval. The task comprises six sub-tasks: identification of event spans, time spans and their attributes, document time relation and the narrative con- tainer relations among events and time expres- sions. We have participated in all six sub- tasks. We have provided with a manually an- notated dataset which comprises of training dataset (293 documents), development dataset (147 documents) and 151 documents as test dataset. We have submitted our work as two systems for the challenge. One system is de- veloped using machine learning techniques, Conditional Random Fields (CRF) and Sup- port Vector machines (SVM) and the other system is developed using deep neural net- work (DNN) techniques. The results show that both systems have given relatively same performance on these tasks. 1 Introduction The interest on extracting temporal information is well versed from the creation of time bank cor- pus (Pustejovsky et al., 2003b) in 2003. A spec- ification language has been developed, TimeML (Markup Language for Temporal and Event Expres- sions) (Pustejovsky et al., 2003a) to conceptualize the events, time expressions and temporal relations using tags (EVENT, TIMEX, TLINKS, ALINKS, SLINKS). Various algorithms have been developed to tag events and time expressions on time bank corpus. Initial works used machine learning algo- rithms with manually extracted features (Mani et al., 2006), syntax and clausal features on inter-sentential events. Later, automated feature selection were used for extracting events and finding the temporal rela- tion(Chambers et al., 2007; Lapata and Lascarides, 2006). TARSQI (Verhagen and Pustejovsky, 2008) is a project employed by the creators of TimeML to develop algorithms for tagging these tags in text. A series of challenges have been organized on TempEval comprising evaluation tasks on events, time expressions and temporal relations on news data. I2b2 2012 (Sun et al., 2013) is the first con- ference to study the temporal information extraction in clinical domain using THYME corpus. It is fol- lowed by Clinical TempEval tasks in semEval2015. Most of the participants of these challenges used CRF and SVM for event extraction with features including the information gathered from different resources like UMLS (Unified Medical Language System), output of TARSQI toolkit, Brown Cluster- ing, Wikipedia and Metamap (Aronson and Lang, 2010). For time expression extraction, various ex- isting tools like SUTIME (Chang and Manning, 2012), HeidelTime (Str¨otgen and Gertz, 2010) and GUTIME were used. And for temporal relation ex- traction various machine learning methods ranging from MaxEnt, Bayesian and SVM to CRF were used incorporating the heuristics and rule based com- ponents. Out of many participants of these chal- lenges/workshops, …

Content truncated to 500 words