Jointly Identifying Predicates, Arguments and Senses using Markov Logic

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 155–163, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Jointly Identifying Predicates, Arguments and Senses using Markov Logic Ivan Meza-Ruiz∗ Sebastian Riedel†‡ ∗School of Informatics, University of Edinburgh, UK †Department of Computer Science, University of Tokyo, Japan ‡Database Center for Life Science, Research Organization of Information and System, Japan ∗I.V.Meza-Ruiz@sms.ed.ac.uk † sebastian.riedel@gmail.com Abstract In this paper we present a Markov Logic Net- work for Semantic Role Labelling that jointly performs predicate identification, frame dis- ambiguation, argument identification and ar- gument classification for all predicates in a sentence. Empirically we find that our ap- proach is competitive: our best model would appear on par with the best entry in the CoNLL 2008 shared task open track, and at the 4th place of the closed track—right be- hind the systems that use significantly better parsers to generate their input features. More- over, we observe that by fully capturing the complete SRL pipeline in a single probabilis- tic model we can achieve significant improve- ments over more isolated systems, in particu- lar for out-of-domain data. Finally, we show that despite the joint approach, our system is still efficient. 1 Introduction Semantic Role Labelling (SRL, M´arquez et al., 2008) is generally understood as the task of iden- tifying and classifying the semantic arguments and modifiers of the predicates mentioned in a sentence. For example, in the case of the following sentence: we are to find out that for the predicate token “plays” with sense “play a role” (play.02) the phrase headed by the token “Haag” is referring to the player (A0) of the play event, and the phrase headed by the token “Elianti” is referring to the role (A1) being played. SRL is considered as a key task for applications that require to answer “Who”, “What”, “Where”, etc. questions, such as Information Extraction, Question Answering and Summarization. Any real-world SRL system needs to make sev- eral decisions, either explicitly or implicitly: which are the predicate tokens of a sentence (predicate identification), which are the tokens that have se- mantic roles with respect to these predicates (argu- ment identification), which are the roles these to- kens play (argument classification), and which is the sense of the predicate (sense disambiguation). In this paper we use Markov Logic (ML), a Statis- tical Relational Learning framework that combines First Order Logic and Markov Networks, to develop a joint probabilistic model over all decisions men- tioned above. The following paragraphs will moti- vate this choice. First, it allows us to readily capture global cor- relations between decisions, such as the constraint that a predicate can only have one agent. This type of correlations has been successfully exploited in several previous SRL approaches (Toutanova et al., 2005; Punyakanok et al., 2005). Second, we can use the joint model to evaluate the benefit of incorporating decisions into the joint model that either have not received much attention within the SRL community (predicate identification and …

Content truncated to 500 words