HeidelTime: High Quality Rule-Based Extraction and Normalization of Temporal Expressions

Proceedings of the 5th International Workshop on Semantic Evaluation, ACL 2010, pages 321–324, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics HeidelTime: High Quality Rule-based Extraction and Normalization of Temporal Expressions Jannik Str¨otgen Institute of Computer Science University of Heidelberg Heidelberg, Germany stroetgen@uni-hd.de Michael Gertz Institute of Computer Science University of Heidelberg Heidelberg, Germany gertz@uni-hd.de Abstract In this paper, we describe HeidelTime, a system for the extraction and normaliza- tion of temporal expressions. HeidelTime is a rule-based system mainly using regu- lar expression patterns for the extraction of temporal expressions and knowledge re- sources as well as linguistic clues for their normalization. In the TempEval-2 chal- lenge, HeidelTime achieved the highest F- Score (86%) for the extraction and the best results in assigning the correct value at- tribute, i.e., in understanding the seman- tics of the temporal expressions. 1 Introduction Temporal annotation of documents, i.e., the ex- traction and chronological ordering of events, is crucial to many NLP applications, e.g., text sum- marization or machine translation. In this paper, we describe our system HeidelTime for the extrac- tion and normalization of temporal expressions in English documents. It was the best-performing system in Task A for English of the TempEval- 2 challenge1. The purpose of this challenge was to evaluate different systems for temporal tagging as well as event and temporal relation extraction since a competitive evaluation helps to drive for- ward research, and temporal annotation is impor- tant for many NLP tasks (Pustejovsky and Verha- gen, 2009). The annotation scheme for tempo- ral expressions, events, and relations is based on TimeML, the ISO standard for temporal annota- tion2. Before using temporal information in other ap- plications is possible, the first task to solve is to ex- tract and normalize temporal expressions (Task A of the challenge, annotated as Timex3). There 1http://semeval2.fbk.eu/ 2http://www.timeml.org/ are two types of approaches to address this prob- lem: rule-based and machine learning ones. We decided to develop a rule-based system since nor- malization can then be supervised in a much eas- ier way. Furthermore, respective systems allow for modular extensions. Although we only participated in Task A, we do not consider the extraction and normalization of temporal expressions in isolation, but use temporal information in combination with other extracted facts, e.g., for the exploration of spatio-temporal information in documents (Str¨otgen et al., 2010). One of our primary objectives is therefore to de- velop a system that can be used in other scenar- ios without any adaptations. Thus, we implement HeidelTime as a UIMA3 (Unstructured Informa- tion Management Architecture) component to in- tegrate the system into our existing document pro- cessing pipeline. Another advantage of our tem- poral tagger is that the user can choose between a precision- and a recall-optimized rule set. In the TempEval-2 challenge, both rule sets achieved top scores in the extraction (F-scores of 86%) and the precision-optimized set achieved the best re- sults for assigning the correct value attributes to the temporal expressions (85% accuracy). The remainder of the paper …

Content truncated to 500 words