LTH: Semantic Structure Extraction using Nonprojective Dependency Trees

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 227–230, Prague, June 2007. c⃝2007 Association for Computational Linguistics LTH: Semantic Structure Extraction using Nonprojective Dependency Trees Richard Johansson and Pierre Nugues Department of Computer Science, Lund University, Sweden {richard, pierre}@cs.lth.se Abstract We describe our contribution to the SemEval task on Frame-Semantic Structure Extrac- tion. Unlike most previous systems de- scribed in literature, ours is based on depen- dency syntax. We also describe a fully auto- matic method to add words to the FrameNet lexical database, which gives an improve- ment in the recall of frame detection. 1 Introduction The existence of links between grammatical rela- tions and various forms of semantic interpretation has long been observed; grammatical relations play a crucial role in theories of linking, i.e. the realiza- tion of the semantic arguments of predicates as syn- tactic units (Manning, 1994; Mel’ˇcuk, 1988). Gram- matical relations may be covered by many defini- tions but it is probably easier to use them as an exten- sion of dependency grammars, where relations take the form of arc labels. In addition, some linguistic phenomena such as wh-movement and discontinu- ous structures are conveniently described using de- pendency syntax by allowing nonprojective depen- dency arcs. It has also been claimed that dependency syntax is easier to understand and to teach to people without a linguistic background. Despite these advantages, dependency syntax has relatively rarely been used in semantic structure ex- traction, with a few exceptions. Ahn et al. (2004) used a post-processing step to convert constituent trees into labeled dependency trees that were then used as input to a semantic role labeler. Pradhan et al. (2005) used a rule-based dependency parser, but the results were significantly worse than when using a constituent parser. This paper describes a system for frame-semantic structure extraction that is based on a dependency parser. The next section presents the dependency grammar that we rely on. We then give the de- tails on the frame detection and disambiguation, the frame element (FE) identification and classification, and dictionary extension, after which the results and conclusions are given. 2 Dependency Parsing with the Penn Treebank The last few years have seen an increasing interest in dependency parsing (Buchholz and Marsi, 2006) with significant improvements of the state of the art, and dependency treebanks are now available for a wide range of languages. The parsing algorithms are comparatively easy to implement and efficient: some of the algorithms parse sentences in linear time (Yamada and Matsumoto, 2003; Nivre et al., 2006). In the semantic structure extraction system, we used the Stanford part-of-speech tagger (Toutanova et al., 2003) to tag the training and test sentences and MaltParser, a statistical dependency parser (Nivre et al., 2006), to parse them. We trained the parser on the Penn Treebank (Mar- cus et al., 1993). The dependency trees used to train the parser were created from the constituent trees using a conversion program (Johansson and Nugues, 2007)1. The converter handles most of the secondary edges in …

Content truncated to 500 words