UTTime: Temporal Relation Classification using Deep Syntactic Features

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 88–92, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics UTTime: Temporal Relation Classification using Deep Syntactic Features Natsuda Laokulrat The University of Tokyo 3-7-1 Hongo, Bunkyo-ku, Tokyo, Japan natsuda@logos.t.u-tokyo.ac.jp Makoto Miwa The University of Manchester 131 Princess Street, Manchester, M1 7DN, UK makoto.miwa@manchester.ac.uk Yoshimasa Tsuruoka The University of Tokyo 3-7-1 Hongo, Bunkyo-ku, Tokyo, Japan tsuruoka@logos.t.u-tokyo.ac.jp Takashi Chikayama The University of Tokyo 3-7-1 Hongo, Bunkyo-ku, Tokyo, Japan chikayama@logos.t.u-tokyo.ac.jp Abstract In this paper, we present a system, UTTime, which we submitted to TempEval-3 for Task C: Annotating temporal relations. The sys- tem uses logistic regression classifiers and ex- ploits features extracted from a deep syntactic parser, including paths between event words in phrase structure trees and their path lengths, and paths between event words in predicate- argument structures and their subgraphs. UT- Time achieved an F1 score of 34.9 based on the graphed-based evaluation for Task C (ranked 2nd) and 56.45 for Task C-relation- only (ranked 1st) in the TempEval-3 evalua- tion. 1 Introduction Temporal annotation is the task of identifying tem- poral relationships between pairs of temporal enti- ties, namely temporal expressions and events, within a piece of text. The temporal relationships are im- portant to support other NLP applications such as textual entailment, document summarization, and question answering. The temporal annotation task consists of several subtasks, including temporal ex- pression extraction, event extraction, and temporal link identification and relation classification. In TempEval-3, there are three subtasks of the temporal annotation process offered, i.e., Task A: Temporal expression extraction and normalization, Task B: Event extraction, and Task C: Annotating temporal relations. This paper presents a system to handle Task C. Based on the annotated data pro- vided, this subtask requires identifying pairs of tem- poral entities and classifying the pairs into one of the 14 relation types according to TimeML (Pustejovsky et al., 2005), i.e., BEFORE, AFTER, IMMEDIATELY BE- FORE, IMMEDIATELY AFTER, INCLUDES, IS INCLUDED, DURING, DURING INVERSE, SIMULTANEOUS, IDENTITY, BEGINS, BEGUN BY, END, and ENDED BY. The motivation behind our work is to utilize syn- tactic and semantic relationships between a pair of temporal entities in the temporal relation classifica- tion task, since we believe that these relationships convey the temporal relation. In addition to general features, which are easily extracted from sentences (e.g., part of speech tags, lemmas, synnonyms), we use features extracted using a deep syntactic parser. The features from the deep parser can be divided into two groups: features from phrase structure trees and features from predicate-argument structures. These features are only applicable in the case that the tem- poral entities appear in the same sentence, so we use only the general features for inter-sentence relations. Predicate-argument structure expresses semantic relations between words. This information can be extracted from a deep syntactic parser. Features from predicate-argument structures can capture im- portant temporal information (e.g., prepositions of time) from sentences effectively. The remaining part of this …

Content truncated to 500 words