Semantic Role Labeling Without Treebanks?

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 192–200, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Semantic Role Labeling Without Treebanks? Stephen A. Boxwell1, Chris Brew2, Jason Baldridge3, Dennis Mehay1, and Sujith Ravi4 1The Ohio State University, {boxwell,mehay}@ling.ohio-state.edu 2The Educational Testing Service, cbrew@ets.org 3The University of Texas at Austin, jbaldrid@mail.utexas.edu 4ISI, sravi@isi.edu Abstract We describe a method for training a se- mantic role labeler for CCG in the absence of gold-standard syntax derivations. Tra- ditionally, semantic role labeling is per- formed by placing human-annotated se- mantic roles on gold-standard syntactic parses, identifying patterns in the syntax- semantics relationship, and then predict- ing roles on novel syntactic analyses. The gold standard syntactic training data can be eliminated from the process by extract- ing training instances from semantic roles projected onto a packed parse chart. This process can be used to rapidly develop NLP tools for resource-poor languages of interest. 1 Introduction Semantic role labeling is the process of generat- ing sets of semantic roles from syntactic analy- ses. The process of training a semantic role la- beler, however, is costly in resources. First, it re- quires gold-standard semantic role data, like Prop- bank (Palmer et al., 2005). Secondly, it requires a detailed syntactic annotation of the same re- source. We are fortunate to have the reasonably- sized Penn Treebank (Marcus et al., 1993) and adaptations for formalisms like Tree Adjoining Grammar (Chen and Shanker, 2004) and Com- binatory Categorial Grammar (Hockenmaier and Steedman, 2007) alongside the Propbank data, but for other languages, such resources are unlikely to be available. There has been work in generating semantic role labelers using gold-standard trees in the absence of semantic training data (F¨urstenau and Lapata, 2009; Lang and Lapata, 2010). But what if we had semantic training data, but no syn- tactic training data? If we could develop and train a semantic role labeler without syntactic training data, we could greatly reduce the cost and devel- opment time of NLP tools for languages of inter- est. One option is to use some automatic means to gen- erate a treebank – instead of creating a corpus of syntax trees by hand, we could use an automatic parser. This, however, leads to a chicken-and-egg problem – we would need a high-quality parse model to choose a single-best analysis for each training sentence, and a parse model needs syn- tactic training data. No automatic parser can cur- rently generate high quality single-best parses in the absence of a parse model. But a parser can, given word tags and combinatory rules, generate a parse forest – a very large collection of possible analyses – and say little or nothing about their rel- ative merit. In this paper, we extract SRL features from the en- tire parse forest, effectively training on every pos- sible parse in the training set simultaneously. This can be done efficiently by representing the parse chart as a hypergraph, enabling us to iterate over every constituent in the parse …

Content truncated to 500 words