ClearTK-TimeML: A minimalist approach to TempEval 2013

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 10–14, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics ClearTK-TimeML: A minimalist approach to TempEval 2013 Steven Bethard Center for Computational Language and Education Research University of Colorado Boulder Boulder, Colorado 80309-0594, USA steven.bethard@colorado.edu Abstract The ClearTK-TimeML submission to Temp- Eval 2013 competed in all English tasks: identi- fying events, identifying times, and identifying temporal relations. The system is a pipeline of machine-learning models, each with a small set of features from a simple morpho-syntactic an- notation pipeline, and where temporal relations are only predicted for a small set of syntac- tic constructions and relation types. ClearTK- TimeML ranked 1st for temporal relation F1, time extent strict F1 and event tense accuracy. 1 Introduction The TempEval shared tasks (Verhagen et al., 2007; Verhagen et al., 2010; UzZaman et al., 2013) have been one of the key venues for researchers to com- pare methods for temporal information extraction. In TempEval 2013, systems are asked to identify events, times and temporal relations in unstructured text. This paper describes the ClearTK-TimeML system submitted to TempEval 2013. This system is based off of the ClearTK framework for machine learning (Ogren et al., 2008)1, and decomposes TempEval 2013 into a series of sub-tasks, each of which is for- mulated as a machine-learning classification problem. The goals of the ClearTK-TimeML approach were: • To use a small set of simple features that can be derived from either tokens, part-of-speech tags or syntactic constituency parses. • To restrict temporal relation classification to a sub- set of constructions and relation types for which the models are most confident. 1http://cleartk.googlecode.com/ Thus, each classifier in the ClearTK-TimeML pipeline uses only the features shared by success- ful models in previous work (Bethard and Martin, 2006; Bethard and Martin, 2007; Llorens et al., 2010; UzZaman and Allen, 2010) that can be derived from a simple morpho-syntactic annotation pipeline2. And each of the temporal relation classifiers is restricted to a particular syntactic construction and to a partic- ular set of temporal relation labels. The following sections describe the models, classifiers and datasets behind the ClearTK-TimeML approach. 2 Time models Time extent identification was modeled as a BIO token-chunking task, where each token in the text is classified as being at the B(eginning) of, I(nside) of, or O(utside) of a time expression. The following features were used to characterize tokens: • The token’s text • The token’s stem • The token’s part-of-speech • The unicode character categories for each character of the token, with repeats merged (e.g. Dec28 would be ‘LuLlNd‘) • The temporal type of each alphanumeric sub-token, derived from a 58-word gazetteer of time words • All of the above features for the preceding 3 and following 3 tokens Time type identification was modeled as a multi- class classification task, where each time is classified 2 OpenNLP sentence segmenter, ClearTK PennTreebank- Tokenizer, Apache Lucene Snowball stemmer, OpenNLP part- of-speech tagger, …

Content truncated to 500 words