KUL: Data-driven Approach to Temporal Parsing of Newswire Articles

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 83–87, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics KUL: A Data-driven Approach to Temporal Parsing of Documents Oleksandr Kolomiyets KU Leuven Celestijnenlaan 200A Heverlee 3001, Belgium Department of Computer Science oleksandr.kolomiyets @cs.kuleuven.be Marie-Francine Moens KU Leuven Celestijnenlaan 200A Heverlee 3001, Belgium Department of Computer Science sien.moens@cs.kuleuven.be Abstract This paper describes a system for temporal processing of text, which participated in the Temporal Evaluations 2013 campaign. The system employs a number of machine learning classifiers to perform the core tasks of: identi- fication of time expressions and events, recog- nition of their attributes, and estimation of temporal links between recognized events and times. The central feature of the proposed sys- tem is temporal parsing – an approach which identifies temporal relation arguments (event- event and event-timex pairs) and the semantic label of the relation as a single decision. 1 Introduction Temporal Evaluations 2013 (TempEval-3) the third iteration of temporal evaluations (afteris TempEval-1 (Verhagen et al., 2007) and TempEval- 2 (Verhagen et al., 2010)) which addresses the task of temporal information processing of text. In contrast to the previous evaluation campaigns where the temporal relation recognition task was simpli- fied by restricting grammatical context (events in adjacent sentences, events and times in the same sentences) and proposed relation pairs, TempEval-3 does not set any context in which temporal re- lations have to be identified. Thus, for temporal relation recognition the challenges consist of: first, detecting a pair of events, or an event and a time that constitutes a temporal relation; and, second, determining what semantic label to assign to the proposed pair. Moreover, TempEval-3 proposes the task of end-to-end temporal processing in which events and times, their attributes and relations have to be identified from a raw text input. In this paper we present a data-driven approach to all-around temporal processing of text. A num- ber of machine-learning detectors were designed to recognize temporal “markables” (events and times) and their attributes. The key feature of our approach is that argument pairs, as well as relations between them, are jointly estimated without specifying in ad- vance the context in which these pairs have to occur. 2 Our Approach 2.1 Timex Processing 2.1.1 Timex Recognition and Normalization The proposed method for timex recognition im- plements a supervised machine learning approach that processes each chunk-phrase derived from the parse tree. Time expressions are detected by the model as phrasal chunks in the parse with their cor- responding spans. In addition, the model is boot- strapped by substitutions of temporal triggers with their synonyms learned by the Latent Words Lan- guage Model (Deschacht et al., 2012) as described in (Kolomiyets et al., 2011). We implemented a logis- tic regression model that makes use of the following features: • the head word of the phrase and its POS tag; • all tokens and POS tags in the phrase as a bag of words; • …

Content truncated to 500 words