Semi-supervised Semantic Role Labeling Using the Latent Words Language Model

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 21–29, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Semi-supervised Semantic Role Labeling Using the Latent Words Language Model Koen Deschacht Department of computer science K.U.Leuven, Belgium koen.deschacht@cs.kuleuven. be Marie-Francine Moens Department of computer science K.U.Leuven, Belgium sien.moens@cs.kuleuven.be Abstract Semantic Role Labeling (SRL) has proved to be a valuable tool for performing auto- matic analysis of natural language texts. Currently however, most systems rely on a large training set, which is manually an- notated, an effort that needs to be repeated whenever different languages or a differ- ent set of semantic roles is used in a cer- tain application. A possible solution for this problem is semi-supervised learning, where a small set of training examples is automatically expanded using unlabeled texts. We present the Latent Words Lan- guage Model, which is a language model that learns word similarities from unla- beled texts. We use these similarities for different semi-supervised SRL methods as additional features or to automatically ex- pand a small training set. We evaluate the methods on the PropBank dataset and find that for small training sizes our best per- forming system achieves an error reduc- tion of 33.27% F1-measure compared to a state-of-the-art supervised baseline. 1 Introduction Automatic analysis of natural language is still a very hard task to perform for a computer. Al- though some successful applications have been de- veloped (see for instance (Chinchor, 1998)), im- plementing an automatic text analysis system is still a labour and time intensive task. Many ap- plications would benefit from an intermediate rep- resentation of texts, where an automatic analysis is already performed which is sufficiently general to be useful in a wide range of applications. Syntactic analysis of texts (such as Part-Of- Speech tagging and syntactic parsing) is an ex- ample of such a generic analysis, and has proved useful in applications ranging from machine trans- lation (Marcu et al., 2006) to text mining in the bio-medical domain (Cohen and Hersh, 2005). A syntactic parse is however a representation that is very closely tied with the surface-form of natural language, in contrast to Semantic Role Labeling (SRL) which adds a layer of predicate-argument information that generalizes across different syn- tactic alternations (Palmer et al., 2005). SRL has received a lot of attention in the research commu- nity, and many systems have been developed (see section 2). Most of these systems rely on a large dataset for training that is manually annotated. In this paper we investigate whether we can develop a system that achieves state-of-the-art semantic role labeling without relying on a large number of la- beled examples. We aim to do so by employing the Latent Words Language Model that learns latent words from a large unlabeled corpus. Latent words are words that (unlike observed words) did not oc- cur at a particular position in a text, but given se- mantic and syntactic constraints from the context could have occurred at that particular position. In …

Content truncated to 500 words