Semantic role labelling with similarity-based generalization using EM-based clustering

Semantic Role Labelling with Similarity-Based Generalization Using EM-based Clustering Ulrike Baldewein, Katrin Erk, Sebastian Padó Saarland University Saarbrücken, Germany {ulrike,erk,pado}@coli.uni-sb.de Detlef Prescher University of Amsterdam Amsterdam, The Netherlands prescher@science.uva.nl Abstract We describe a system for semantic role assignment built as part of the Senseval III task, based on an off-the-shelf parser and Maxent and Memory-Based learners. We focus on generalisation using several similarity measures to increase the amount of train- ing data available and on the use of EM-based clus- tering to improve role assignment. Our final score is Precision=73.6%, Recall=59.4% (F=65.7). 1 Introduction This paper describes a study in semantic role la- belling in the context of the Senseval III task, for which the training and test data were both drawn from the current FrameNet release (Johnson et al., 2002). We concentrated on two questions: first, whether role assignment can be improved by gener- alisation over training instances using different sim- ilarity measures; and second, the impact of EM- based clustering, both in deriving more informative selectional preference features and in the generali- sations mentioned above. The basis of our experi- ments was formed by off-the-shelf statistical tools for data processing and modelling. After listing our data preparation steps (Sec. 2) and features (Sec. 3), we describe our classification procedure and the learners we used (Sec. 4). Sec. 5 outlines our experiments in similarity-based gener- alisations, and Section 6 discusses our results. 2 Data and Instances Parsing. To tag and parse the data, we used LoPar (Schmid, 2000), a probabilistic context- free parser, which comes with a Head-Lexicalised Grammar for English (Carroll and Rooth, 1998). We considered only the most probable parse for each sentence and simplified parse trees by elim- inating unary nodes. The resulting nodes form the instances of our classification. We used the Stuttgart TreeTagger (Schmid, 1994) to lemmatise constituent heads. Projection of role labels. FrameNet provides se- mantic roles as character offsets. We labelled those instances (i.e. nodes in the parse tree) with gold standard semantic roles which corresponded to roles’ maximal projections. 13.95% of roles in the training corpus spanned more than one parse tree node. Figure 1 shows an example sentence for the AWARENESS frame. The nodes’ respective seman- tic role labels are given in small caps, and the target predicate is marked in boldface. S (NONE) NP (COGNIZER) Peter VP (NONE) V (NONE) does not VP (NONE) know NP (CONTENT) the answer Figure 1: Example parse tree with role labels Semantic clustering. We used clustering to gen- eralise over possible fillers of roles. In a first model, we derived a probability distribution ���� for pairs ��� �� �� , where � is a target:role combination and �  is the head lemma of a role filler. The key idea is that � and �  are mutually independent, but conditioned on an unobserved class  . In this manner, we define the probability of ��� � �  �      as: �� � � � ! #" �� …

Content truncated to 500 words