Semantic Role Labeling Using Lexical Statistical Information

Proceedings of the 9th Conference on Computational Natural Language Learning (CoNLL), pages 213–216, Ann Arbor, June 2005. c⃝2005 Association for Computational Linguistics Semantic Role Labeling Using Lexical Statistical Information Simone Paolo Ponzetto and Michael Strube EML Research gGmbH Schloss-Wolfsbrunnenweg 33 69118 Heidelberg, Germany http://www.eml-research.de/nlp/ Abstract Our system for semantic role labeling is multi-stage in nature, being based on tree pruning techniques, statistical methods for lexicalised feature encoding, and a C4.5 decision tree classifier. We use both shal- low and deep syntactic information from automatically generated chunks and parse trees, and develop a model for learning the semantic arguments of predicates as a multi-class decision problem. We evalu- ate the performance on a set of relatively ‘cheap’ features and report an F1 score of 68.13% on the overall test set. 1 Introduction This paper presents a system for the CoNLL 2005 Semantic Role Labeling shared task (Carreras & M`arquez, 2005), which is based on the current re- lease of the English PropBank data (Palmer et al., 2005). For the 2005 edition of the shared task are available both syntactic and semantic information. Accordingly, we make use of both clausal, chunk and deep syntactic (tree structure) features, named entity information, as well as statistical representa- tions for lexical item encoding. The set of features and their encoding reflect the necessity of limiting the complexity and dimension- ality of the input space. They also provide the clas- sifier with enough information. We explore here the use of a minimal set of compact features for seman- tic role prediction, and show that a feature-based statistical encoding of lexicalised features such as predicates, head words, local contexts and PoS by means of probability distributions provides an effi- cient way of representing the data, with the feature vectors having a small dimensionality and allowing to abstract from single words. 2 System description 2.1 Preprocessing During preprocessing the predicates’ semantic argu- ments are mapped to the nodes in the parse trees, a set of hand-crafted shallow tree pruning rules are ap- plied, probability distributions for feature represen- tation are generated from training data1, and feature vectors are extracted. Those are finally fed into the classifier for semantic role classification. 2.1.1 Tree node mapping of semantic arguments and named entities Following Gildea & Jurafsky (2002), (i) labels matching more than one constituent due to non- branching nodes are taken as labels of higher con- stituents, (ii) in cases of labels with no correspond- ing parse constituent, these are assigned to the par- tial match given by the constituent spanning the shortest portion of the sentence beginning at the la- bel’s span left boundary and lying entirely within it. We drop the role or named entity label if such suit- able constituent could not be found2. 1All other processing steps assume a uniform treatment of both training and test data. 2The percentage of roles for which no valid tree node could be found amounts to 3% for the training and 7% for the devel- opment set. These results are …

Content truncated to 500 words