JU_CSE: A CRF Based Approach to Annotation of Temporal Expression, Event and Temporal Relations

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 64–72, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics JU_CSE: A CRF Based Approach to Annotation of Temporal Expres- sion, Event and Temporal Relations Anup Kumar Kolya1, Amitava Kundu1, Rajdeep Gupta1 Asif Ekbal2, Sivaji Bandyopadhyay1 1Dept. of Computer Science & Engineering 2Dept. of Computer Science & Engineering Jadavpur Univeristy IIT Patna Kolkata-700 032, India Patna-800 013, India {anup.kolya,amitava.jucse, rajdeepgupta20}@gmail.com asif@iitp.ac.in, sivaji_ju_cse@yahoo.com Abstract In this paper, we present the JUCSE system, designed for the TempEval-3 shared task. The system extracts events and temporal infor- mation from natural text in English. We have participated in all the tasks of TempEval-3, namely Task A, Task B & Task C. We have primarily utilized the Conditional Random Field (CRF) based machine learning tech- nique, for all the above tasks. Our system seems to perform quite competitively in Task A and Task B. In Task C, the system’s per- formance is comparatively modest at the ini- tial stages of system development. We have incorporated various features based on differ- ent lexical, syntactic and semantic infor- mation, using Stanford CoreNLP and Wordnet based tools. 1 Introduction Temporal information extraction has been a popu- lar and interesting research area of Natural Lan- guage Processing (NLP) for quite some time. Generally, a lot of events are described in a variety of newspaper texts, stories and other important documents where the different events described happen at different time instants. The temporal location and ordering of these events are either specified or implied. Automatic identification of time expressions and events and annotation of temporal relations constitute an important task in text analysis. These are also important in a wide range of NLP applications that include temporal question answering, machine translation and doc- ument summarization. A lot of research in the area of temporal infor- mation extraction has been conducted on multiple languages, including English and several European languages. The TimeML was first developed in 2002 in an extended workshop called TERQAS (Time and Event Recognition for Question An- swering Systems) and, in 2003, it was further de- veloped in the context of the TANGO workshop (TimeML Annotation Graphical Organizer). Since then most of the works in this research arena have been conducted in English. The variety of works include TimeML (Pustejovsky et al., 2003), the development of a temporally annotated corpus Time-Bank (Pustejovsky et al., 2003), the temporal evaluation challenges TempEval-1 (Verhagen et al., 2007), TempEval-2 (Pustejovsky and Verha- gen, 2010). In the series of Message Understanding Conferences (MUCs) that started from 1987 and the Sheffield Temporal Annotation scheme (STAG) (Setzer &Gaizauskas, 2000) the aim was to identify events in news text and determine their relationship with points on a temporal line. In the series of TempEval evaluation exercises, TempEval-1 was the first one where the focus was on identification of three types of temporal rela- tion: relation between an event and a time expres- sion in the …

Content truncated to 500 words