Temporal Information Processing of a New Language: Fast Porting with Minimal Resources

Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 671–677, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Temporal information processing of a new language: fast porting with minimal resources Francisco Costa and Ant´onio Branco Universidade de Lisboa Abstract We describe the semi-automatic adapta- tion of a TimeML annotated corpus from English to Portuguese, a language for which TimeML annotated data was not available yet. In order to validate this adaptation, we use the obtained data to replicate some results in the literature that used the original English data. The fact that comparable results are obtained indi- cates that our approach can be used suc- cessfully to rapidly create semantically an- notated resources for new languages. 1 Introduction Temporal information processing is a topic of nat- ural language processing boosted by recent eval- uation campaigns like TERN2004,1 TempEval-1 (Verhagen et al., 2007) and the forthcoming TempEval-22 (Pustejovsky and Verhagen, 2009). For instance, in the TempEval-1 competition, three tasks were proposed: a) identifying the temporal relation (such as overlap, before or after) hold- ing between events and temporal entities such as dates, times and temporal durations denoted by ex- pressions (i.e. temporal expressions) occurring in the same sentence; b) identifying the temporal re- lation holding between events expressed in a doc- ument and its creation time; c) identifying the tem- poral relation between the main events expressed by two adjacent sentences. Supervised machine learning approaches are pervasive in the tasks of temporal information pro- cessing. Even when the best performing sys- tems in these competitions are symbolic, there are machine learning solutions with results close to their performance. In TempEval-1, where there were statistical and rule-based systems, almost 1http://timex2.mitre.org 2http://www.timeml.org/tempeval2 all systems achieved quite similar results. In the TERN2004 competition (aimed at identifying and normalizing temporal expressions), a symbolic system performed best, but since then machine learning solutions, such as (Ahn et al., 2007), have appeared that obtain similar results. These evaluations made available sets of anno- tated data for English and other languages, used for training and evaluation. One natural question to ask is whether it is feasible to adapt the training and test data made available in these competitions to other languages, for which no such data still ex- ist. Since the annotations are largely of a seman- tic nature, not many changes need to be done in the annotations once the textual material is trans- lated. In essence, this would be a fast way to create temporal information processing systems for lan- guages for which there are no annotated data yet. In this paper, we report on an experiment that consisted in adapting the English data of TempEval-1 to Portuguese. The results of ma- chine learning algorithms over the data thus ob- tained are compared to those reported for the En- glish TempEval-1 competition. Since the results are quite similar, this permits to conclude that such an approach can rapidly generate relevant and comparable data and is useful when porting tem- poral …

Content truncated to 500 words