Projection-based Acquisition of a Temporal Labeller

Projection-based Acquisition of a Temporal Labeller Kathrin Spreyer∗ Department of Linguistics University of Potsdam Germany spreyer@uni-potsdam.de Anette Frank Dept. of Computational Linguistics University of Heidelberg Germany frank@cl.uni-heidelberg.de Abstract We present a cross-lingual projection frame- work for temporal annotations. Auto- matically obtained TimeML annotations in the English portion of a parallel corpus are transferred to the German translation along a word alignment. Direct projection augmented with shallow heuristic knowl- edge outperforms the uninformed baseline by 6.64% F1-measure for events, and by 17.93% for time expressions. Subsequent training of statistical classifiers on the (im- perfect) projected annotations significantly boosts precision by up to 31% to 83.95% and 89.52%, respectively. 1 Introduction In recent years, supervised machine learning has be- come the standard approach to obtain robust and wide-coverage NLP tools. But manually annotated training data is a scarce and expensive resource. An- notation projection (Yarowsky and Ngai, 2001) aims at overcoming this resource bottleneck by scaling conceptually monolingual resources and tools to a multilingual level: annotations in existing monolin- gual corpora are transferred to a different language along the word alignment to a parallel corpus. In this paper, we present a projection framework for temporal annotations. The TimeML specifica- tion language (Pustejovsky et al., 2003a) defines an annotation scheme for time expressions (timex for ∗ The first author was affiliated with Saarland University (Saarbr¨ucken, Germany) at the time of writing. John [met]event Mary [last night]timex. John [traf]event Mary [gestern Abend]timex. Figure 1: Annotation projection. short) and events, and there are tools for the auto- matic TimeML annotation of English text (Verha- gen et al., 2005). Similar rule-based systems exist for Spanish and Italian (Saquete et al., 2006). How- ever, such resources are restricted to a handful of languages. We employ the existing TimeML labellers to an- notate the English portion of a parallel corpus, and automatically project the annotations to the word- aligned German translation. Fig. 1 shows a simple example. The English sentence contains an event and a timex annotation. The event-denoting verb met is aligned with the German traf, hence the latter also receives the event tag. Likewise, the components of the multi-word timex last night align with German gestern and abend, respectively, and the timex tag is transferred to the expression gestern abend. Projection-based approaches to multilingual an- notation have proven adequate in various domains, including part-of-speech tagging (Yarowsky and Ngai, 2001), NP-bracketing (Yarowsky et al., 2001), dependency analysis (Hwa et al., 2005), and role se- mantic analysis (Pad´o and Lapata, 2006). To our knowledge, the present proposal is the first to apply projection algorithms to temporal annotations. 489 Cross-lingually projected information is typically noisy, due to errors in the source annotations as well as in the word alignment. Moreover, success- ful projection relies on the direct correspondence assumption (DCA, Hwa et al. (2002)) which de- mands that the annotations in the source text be homomorphous with those in its (literal) transla- tion. The DCA has been found to hold, to a sub- stantial degree, for the above mentioned …

Content truncated to 500 words