Bootstrapping Semantic Role Labelers from Parallel Data

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 1: Proceedings of the Main Conference and the Shared Task, pages 317–327, Atlanta, Georgia, June 13-14, 2013. c⃝2013 Association for Computational Linguistics Bootstrapping Semantic Role Labelers from Parallel Data Mikhail Kozhevnikov Ivan Titov Saarland University, Postfach 15 11 50 66041 Saarbr¨ucken, Germany {mkozhevn|titov}@mmci.uni-saarland.de Abstract We present an approach which uses the sim- ilarity in semantic structure of bilingual par- allel sentences to bootstrap a pair of seman- tic role labeling (SRL) models. The setting is similar to co-training, except for the inter- mediate model required to convert the SRL structure between the two annotation schemes used for different languages. Our approach can facilitate the construction of SRL models for resource-poor languages, while preserving the annotation schemes designed for the tar- get language and making use of the limited re- sources available for it. We evaluate the model on four language pairs, English vs German, Spanish, Czech and Chinese. Consistent im- provements are observed over the self-training baseline. 1 Introduction The success of statistical modeling methods in a va- riety of natural language processing (NLP) tasks in the last decade depended crucially on the availability of annotated resources for their training. And while sizable resources for most standard tasks are only available for a few languages, the human effort re- quired to achieve reasonable performance on such tasks for other languages may be significantly re- duced by leveraging existing resources and the sim- ilarities between languages. This idea has lead to the development of cross- lingual annotation projection approaches, which make use of parallel corpora (Pad´o and Lapata, 2009), as well as attempts to adapt models directly to other languages (McDonald et al., 2011). In this paper we consider correspondences between SRL structures in translated sentences from a different perspective. Most cross-lingual annotation projec- tion approaches transfer the source language anno- tation scheme to the target language without modifi- cation, which makes it hard to combine their output with existing target language resources, as annota- tion schemes may vary significantly. We instead ad- dress the problem of information transfer between two existing annotation schemes (figure 1) for a pair of languages using an intermediate model of role correspondence (RCM). An RCM models the prob- ability of a pair of corresponding arguments being assigned a certain pair of roles. We then use it to guide a pair of monolingual models toward compat- ible predictions on parallel data in order to extend the coverage and/or accuracy of one or both models. Romanian is not taught in their schools . Ve školách se neučí rumunsky . A1 PAT AM-LOC LOC AM-NEG Figure 1: Role correspondence in parallel sentences, an example. The notion of compatibility here is highly non- trivial, even for sentences translated as close to the original as possible. Zhuang and Zong (2010), for example, observe that in the English-Chinese paral- lel PropBank (Palmer et al., 2005b) corresponding arguments often bear different labels, even though the same inventory of semantic roles is used …

Content truncated to 500 words