Semantic Role Labelling with minimal resources: Experiments with French

Proceedings of the Third Joint Conference on Lexical and Computational Semantics (*SEM 2014), pages 87–92, Dublin, Ireland, August 23-24 2014. Semantic Role Labelling with minimal resources: Experiments with French Rasoul Kaljahi†‡, Jennifer Foster†, Johann Roturier‡ †NCLT, School of Computing, Dublin City University, Ireland {rkaljahi, jfoster}@computing.dcu.ie ‡Symantec Research Labs, Dublin, Ireland johann roturier@symantec.com Abstract This paper describes a series of French se- mantic role labelling experiments which show that a small set of manually anno- tated training data is superior to a much larger set containing semantic role labels which have been projected from a source language via word alignment. Using uni- versal part-of-speech tags and dependen- cies makes little difference over the orig- inal fine-grained tagset and dependency scheme. Moreover, there seems to be no improvement gained from projecting se- mantic roles between direct translations than between indirect translations. 1 Introduction Semantic role labelling (SRL) (Gildea and Juraf- sky, 2002) is the task of identifying the predicates in a sentence, their semantic arguments and the roles these arguments take. The last decade has seen considerable attention paid to statistical SRL, thanks to the existence of two major hand-crafted resources for English, namely, FrameNet (Baker et al., 1998) and PropBank (Palmer et al., 2005). Apart from English, only a few languages have SRL resources and these resources tend to be of limited size compared to the English datasets. French is one of those languages which suffer from a scarcity of hand-crafted SRL resources. The only available gold-standard resource is a small set of 1000 sentences taken from Europarl (Koehn, 2005) and manually annotated with Prop- bank verb predicates (van der Plas et al., 2010b). This dataset is then used by van der Plas et al. (2011) to evaluate their approach to projecting the SRLs of English sentences to their translations This work is licensed under a Creative Commons Attribution 4.0 International Licence. Page numbers and proceedings footer are added by the organisers. Licence details: http: //creativecommons.org/licenses/by/4.0/ in French. They additionally build a large, “ar- tificial” or automatically labelled dataset of ap- proximately 1M Europarl sentences by projecting the SRLs from English sentences to their French translations and use it for training an SRL system. We build on the work of van der Plas et al. (2010b) by answering the following questions: 1) How much artificial data is needed to train an SRL system? 2) Is it better to use direct trans- lations than indirect translations, i.e. is it better to use for projection a source-target pair where the source represents the original sentence and the target represents its direct translation as opposed to a source-target pair where the source and tar- get are both translations of an original sentence in a third language? 3) Is it better to use coarse- grained syntactic information (in the form of uni- versal part-of-speech tags and universal syntactic dependencies) than to use fine-grained syntactic information? We find that SRL performance lev- els off after only 5K training sentences obtained via projection and that direct translations …

Content truncated to 500 words