Semantic Role Labelling with Tree Conditional Random Fields

Proceedings of the 9th Conference on Computational Natural Language Learning (CoNLL), pages 169–172, Ann Arbor, June 2005. c⃝2005 Association for Computational Linguistics Semantic Role Labelling with Tree Conditional Random Fields Trevor Cohn and Philip Blunsom University of Melbourne, Australia tacohn@csse.unimelb.edu.au and pcbl@csse.unimelb.ed.au Abstract In this paper we apply conditional random fields (CRFs) to the semantic role labelling task. We define a random field over the structure of each sentence’s syntactic parse tree. For each node of the tree, the model must predict a semantic role label, which is interpreted as the labelling for the corresponding syntactic constituent. We show how modelling the task as a tree labelling problem allows for the use of efficient CRF inference algorithms, while also increasing generalisation performance when compared to the maximum entropy classifier.equivalent We have participated in the CoNLL-2005 shared task closed challenge with full syntactic information. 1 Introduction The semantic role labelling task (SRL) involves identifying which groups of words act as arguments to a given predicate. These arguments must be labelled with their role with respect to the predicate, indicating how the proposition should be semantically interpreted. We apply conditional random fields (CRFs) to the task of SRL proposed by the CoNLL shared task 2005 (Carreras and M`arquez, 2005). CRFs are undirected graphical models which define a condi- tional distribution over labellings given an obser- vation (Lafferty et al., 2001). These models allow for the use of very large sets of arbitrary, over- lapping and non-independent features. CRFs have been applied with impressive empirical results to the tasks of named entity recognition (McCallum and Li, 2003; Cohn et al., 2005), part-of-speech (PoS) tagging (Lafferty et al., 2001), noun phrase chunk- ing (Sha and Pereira, 2003) and extraction of table data (Pinto et al., 2003), among other tasks. While CRFs have not been used to date for SRL, their close cousin, the maximum entropy model has been, with strong generalisation performance (Xue and Palmer, 2004; Lim et al., 2004). Most CRF implementations have been specialised to work with chain structures, where the labels and observations form a linear sequence. Framing SRL as a linear tagging task is awkward, as there is no easy model of adjacency between the candidate constituent phrases. Our approach simultaneously performs both con- stituent selection and labelling, by defining an undi- rected random field over the parse tree. This allows the modelling of interactions between parent and child constituents, and the prediction of an optimal argument labelling for all constituents in one pass. The parse tree forms an acyclic graph, meaning that efficient exact inference in a CRF is possible using belief propagation. 2 Data The data used for this task was taken from the Propbank corpus, which supplements the Penn Treebank with semantic role annotation. Full details of the data set are provided in Carreras and M`arquez (2005). 2.1 Data Representation From each training instance we derived a tree, using the parse structure from the Collins parser. The 169 nodes in the trees were relabelled with a semantic …

Content truncated to 500 words