Improving Semantic Role Labeling with Word Sense

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 246–249, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Improving Semantic Role Labeling with Word Sense Wanxiang Che, Ting Liu and Yongqiang Li Research Center for Information Retrieval MOE-Microsoft Key Laboratory of Natural Language Processing and Speech School of Computer Science and Technology Harbin Institute of Technology, China, 150001 {car, tliu, yqli}@ir.hit.edu.cn Abstract Semantic role labeling (SRL) not only needs lexical and syntactic information, but also needs word sense information. However, be- cause of the lack of corpus annotated with both word senses and semantic roles, there is few research on using word sense for SRL. The release of OntoNotes provides an oppor- tunity for us to study how to use word sense for SRL. In this paper, we present some novel word sense features for SRL and find that they can improve the performance significantly. 1 Introduction Semantic role labeling (SRL) is a kind of shallow sentence-level semantic analysis and is becoming a hot task in natural language processing. SRL aims at identifying the relations between the predicates in a sentence and their associated arguments. At present, the main stream researches are focusing on feature engineering or combination of multiple results. Word senses are important information for rec- ognizing semantic roles. For example, if we know “cat” is an “agent” of the predicate “eat” in a sentence, we can guess that “dog” can also be an “agent” of “eat”. Word sense has been suc- cessfully used in many natural language process- ing tasks, such as machine translation (Chan et al., 2007; Carpuat and Wu, 2007). CoNLL 2008 shared task (Surdeanu et al., 2008) first introduced the pred- icate classification task, which can be regarded as the predicate sense disambiguation. Meza-Ruiz and Riedel (2009) has shown that the predicate sense can improve the final SRL performance. However, there is few discussion about the concrete influence of all word senses, i.e. the words besides predicates. The major reason is lacking the corpus, which is both an- notated with all word senses and semantic roles. The release of OntoNotes corpus provides an op- portunity for us to verify whether all word senses can help SRL. OntoNotes is a large corpus annotated with constituency trees (based on Penn Treebank), predicate argument structures (based on Penn Prop- Bank) and word senses. It has been used in some natural language processing tasks, such as joint pars- ing and named entity recognition (Finkel and Man- ning, 2009) and word sense disambiguation (Zhong et al., 2008). In this paper, we regard the word sense informa- tion as additional SRL features. We compare three categories of word sense features (subtree-word re- lated sense, predicate sense, and sense path) and find that the subtree-word related sense feature is ineffec- tive, however, the predicate sense and the sense path features can improve the SRL performance signifi- cantly. 2 Data Preparation In our experiments, we use the OntoNotes Release 2.01 corpus (Hovy et …

Content truncated to 500 words