A Supervised Machine Learning Approach for Event-Event Relation Identification

A Supervised Machine Learning Approach for Temporal Information Extraction Anup Kumar Kolya a, Asif Ekbal b, and Sivaji Bandyopadhyay c a, cDepartment of Computer Science and Engineering, Jadavpur University, Kolkata-700032, India anup.kolya@gmail.com a, sivaji_cse_ju@yahoo.com c bDepartment of Information Engineering and Computer Science, University of Trento, Italy asif.ekbal@gmail.com b Abstract. Temporal information extraction is an interesting research area in Natural Language Processing (NLP). Here, the main task involves identification of the different relations between various events and time expressions in a document. The relations are then classified into some predefined categories like BEFORE, AFTER, OVERLAP, BEFORE-OR-OVERLAP, OVERLAP-OR-AFTER and VAGUE. In this paper, we report our works of temporal information extraction along the lines of TempEval-2007 evaluation challenge. We adapt supervised machine learning approach for solving the problems of all the three tasks, namely A, B and C. Initially, a baseline system is developed by considering the most frequent temporal relation in the corresponding task’s training data. Evaluation results on the TempEval-2007 datasets yield the F-score values of 59.8%, 73.8% and 43.8% for Tasks A, B and C, respectively under the strict evaluation scheme. All these systems show the F-score values of 61.1%, 74.8% and 46.9% for Tasks A, B and C, respectively under the relaxed evaluation scheme. For the sub-ordinate event in Task C, the system shows the F-score values of 55.1% and 56.9% under the strict and relaxed evaluation scheme, respectively. Keywords: Temporal Relation Identification, TimeML, Conditional Random Field, TempEval-2007, Tasks A, B and C. 1 Introduction Temporal information extraction has become a popular and interesting research area of Natural Language Processing (NLP) since the last few years. Generally, events are described in different newspaper texts, stories and other important documents where events happen in time. Many text processing applications require identifying these events described in a text and locating these in time. This is also important in a wide range of NLP applications that include temporal question answering, machine translation and document summarization etc. The TempEval-2007 challenge (Verhagen et al., 2007) addressed this question by establishing a common corpus on which various research teams came up with different approaches to find temporal relations. In TempEval-2007, the following types of temporal relations (i.e. event-event and/or event-time) were considered: • Task A: Relation between the events and times within the same sentence. • Task B: Relation between events and document creation times. • Task C. Relation between the verb events in adjacent sentences. In each of these tasks, systems attempted to annotate appropriate pairs with one of the relations, namely BEFORE, BEFORE-OR-OVERLAP, OVERLAP, OVERLAP-OR-AFTER, AFTER or VAGUE. The participating teams were instructed to find all temporal relations of these types in a corpus of newswire documents. In the literature, temporal relation identification has been treated as a classification problem and solved using machine learning in a number of proposals (Boguraev et al., 2005; Mani et al., 2007; Mani et al., 2007; Chambers et al., 2007). Some of the TempEval-2007 participants PACLIC 24 Proceedings 447 (Verhagen et al., 2007) also proposed several …

Content truncated to 500 words