Collective Semantic Role Labelling with Markov Logic

CoNLL 2008: Proceedings of the 12th Conference on Computational Natural Language Learning, pages 193–197 Manchester, August 2008 Collective Semantic Role Labelling with Markov Logic Sebastian Riedel Ivan Meza-Ruiz Institute for Communicating and Collaborative Systems School of Informatics University of Edinburgh, Scotland {S.R.Riedel,I.V.Meza-Ruiz}@sms.ed.ac.uk Abstract This paper presents our system for the Open Track of the CoNLL 2008 Shared Task (Surdeanu et al., 2008) in Joint De- pendency Parsing1 and Semantic Role La- belling. We use Markov Logic to define a joint SRL model and achieve a semantic F-score of 74.59%, the second best in the Open Track. 1 Introduction Many SRL systems use a two-stage pipeline that first extracts possible argument candidates (argu- ment identification) and then assigns argument labels to these candidates (argument classifica- tion) (Xue and Palmer, 2004). If we also con- sider the necessary previous step of identifying the predicates and their senses (predicate identi- fication) this yields a three-stage pipeline: predi- cate identification, argument identification and ar- gument classification. Our system, on the other hand, follows a joint approach in the spirit of Toutanova et al. (2005) and performs the above steps collectively . We de- cided to use Markov Logic (ML, Richardson and Domingos, 2005), a First Order Probabilistic Lan- guage, to develop a global probabilistic model of SRL. By using ML we are able to incorporate the dependencies between the decisions of different stages in the pipeline and the well-known global correlations between the arguments of a predi- cate (Punyakanok et al., 2005). And since learning c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. 1Note that in this work we do not consider the parsing task; instead we use the provided dependencies of the open track datatsets. and inference methods were already implemented in the ML software we use, only minimal engineer- ing efforts had to be done. In contrast to the work of Toutanova et al. (2005) our system applies online learning to train its pa- rameters and exact inference to predict a collective role labelling. Moreover, we jointly label the argu- ments of all predicates in a sentence. This allows us, for example, to require that certain tokens have to be an argument of some predicates in the sen- tence. In this paper we also investigate the impact of different levels of interaction between the layers of the joint SRL model. We find that a probabilis- tic model which resembles a traditional bottom-up pipeline (though jointly trained and globally nor- malised) performs better than the complete joint model on the WSJ test set and worse on the Brown test set. The worst performance is observed when no interaction between SRL stages is allowed. In terms of semantic F-score (74.59%) our sub- mitted results are the second best in the Open Track of the Shared Task. Our error analysis in- dicates that a) the training regime can be improved and b) nominalizations are difficult to handle for the model as it is. In the next …

Content truncated to 500 words