Using Cross-Lingual Projections to Generate Semantic Role Labeled Annotated Corpus for Urdu - A Resource Poor Language

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 797–805, Beijing, August 2010 Using Cross-Lingual Projections to Generate Semantic Role Labeled Corpus for Urdu - A Resource Poor Language Smruthi Mukund CEDAR University at Buffalo smukund@buffalo.edu Debanjan Ghosh Thomson Reuters R&D debanjan.ghosh@ thomsonreuters.com Rohini K. Srihari CEDAR University at Buffalo rohini@cedar.buffalo.edu Abstract In this paper we explore the possibility of using cross lingual projections that help to automatically induce role-semantic annotations in the PropBank paradigm for Urdu, a resource poor language. This technique provides annotation projections based on word alignments. It is relatively inexpensive and has the potential to re- duce human effort involved in creating semantic role resources. The projection model exploits lexical as well as syntac- tic information on an English-Urdu paral- lel corpus. We show that our method ge- nerates reasonably good annotations with an accuracy of 92% on short structured sentences. Using the automatically gen- erated annotated corpus, we conduct pre- liminary experiments to create a semantic role labeler for Urdu. The results of the labeler though modest, are promising and indicate the potential of our technique to generate large scale annotations for Urdu. 1 Introduction Semantic Roles (also known as thematic roles) help to understand the semantic structure of a document (Fillmore, 1968). At a fundamental level, they help to capture the similarities and differences in the meaning of verbs via the ar- guments they define by generalizing over surface syntactic configurations. In turn, these roles aid in domain independent understanding as the se- mantic frames and semantic understanding sys- tems do not depend on the syntactic configura- tion for each new application domain. Identify- ing semantic roles benefit several language processing tasks - information extraction (Sur- deanu et al., 2003), text categorization (Moschitti, 2008) and finding relations in textual entailment (Burchardt and Frank 2006). Automatically identifying semantic roles is of- ten referred to as shallow semantic parsing (Gil- dea and Jurafsky, 2002). For English, this process is facilitated by the existence of two main SRL annotated corpora – FrameNet (Baker et al., 1998) and PropBank (Palmer et al., 2005). Both datasets mark almost all surface realizations of semantic roles. FrameNet has 800 semantic frames that cover 120,000 example sentences1. PropBank has annotations that cover over 113,000 predicate-argument structures. Clearly English is well supported with resources for se- mantic roles. However, there are other widely spoken resource poor languages that are not as privileged. The PropBank based resources avail- able for languages like Chinese (Xue and Palmer, 2009), Korean (Palmer et al., 2006) and Spanish (Taule, 2008) are only about two-thirds the size of the English PropBank. Several alternative techniques have been ex- plored in the literature to generate semantic role labeled corpora for resource poor languages as providing manually annotated data is time con- suming and involves intense human labor. Am- bati and Chen (2007) have conducted an exten- sive survey and outlined the benefits of using parallel corpora to transfer annotations. A wide range of annotations from part of speech …

Content truncated to 500 words