Non-atomic Classification to Improve a Semantic Role Labeler for a Low-resource Language

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 95–99, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Non-atomic Classification to Improve a Semantic Role Labeler for a Low-resource Language Richard Johansson Spr˚akbanken, Department of Swedish, University of Gothenburg Box 100, SE-40530 Gothenburg, Sweden richard.johansson@gu.se Abstract Semantic role classification accuracy for most languages other than English is constrained by the small amount of annotated data. In this pa- per, we demonstrate how the frame-to-frame relations described in the FrameNet ontology can be used to improve the performance of a FrameNet-based semantic role classifier for Swedish, a low-resource language. In order to make use of the FrameNet relations, we cast the semantic role classification task as a non-atomic label prediction task. The ex- periments show that the cross-frame general- ization methods lead to a 27% reduction in the number of errors made by the classifier. For previously unseen frames, the reduction is even more significant: 50%. 1 Introduction The FrameNet lexical database and annotated cor- pora, based on the theory of semantic frames (Fill- more et al., 2003), have allowed the implementa- tion of automatic systems to extract semantic roles (Gildea and Jurafsky, 2002; Johansson and Nugues, 2007; M`arquez et al., 2008; Das et al., 2010). Since the original FrameNet is developed for the English language, most research on semantic role extraction has focused exclusively on English. How- ever, the English FrameNet has inspired similar ef- forts for other languages. For instance, the ongo- ing development of a Swedish FrameNet (Borin et al., 2010) allows us to investigate the feasibility of using this resource in constructing an automatic role-semantic analyzer for Swedish. However, due to the fact that the Swedish FrameNet annotation process is in a fairly early stage, not much anno- tated material is available, and this limits the perfor- mance attainable by automatic classifiers trained on these data. In particular, the scarce amount of data makes it very hard for the machine learning meth- ods to discern general linguistic facts concerning the syntactic–semantic linking patterns, such as the rela- tion between the voice of a verb, the syntactic func- tions of its arguments, and the semantic roles of the arguments (Dowty, 1991). In this paper, we show that the inter-frame rela- tions described in the FrameNet ontology allow us to generalize across frames. This allows the clas- sifier to learn general linguistic facts, and it also leads to more efficient use of the annotated data. To allow this kind of generalization, we formulate the semantic role selection problem as a classifica- tion task with non-atomic labels. This cross-frame generalization method reduces the number of errors made by the classifier by 27%, improving the accu- racy from 54.4 to 66.5. When evaluating on frames for which the classifier has not been trained, the ac- curacy improves from 7.2 (random performance) to 53.4, a 50% error reduction. 2 The Swedish FrameNet The Swedish FrameNet, SweFN, is a lexical re- source under development (Friberg Heppin and Toporowska Gronostaj, 2012), …

Content truncated to 500 words