Joint Learning of Preposition Senses and Semantic Roles of Prepositional Phrases

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 450–458, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Joint Learning of Preposition Senses and Semantic Roles of Prepositional Phrases Daniel Dahlmeier1, Hwee Tou Ng1,2, Tanja Schultz3 1NUS Graduate School for Integrative Sciences and Engineering 2Department of Computer Science, National University of Singapore 3Cognitive Systems Lab, University of Karlsruhe {danielhe,nght}@comp.nus.edu.sg tanja@ira.uka.de Abstract The sense of a preposition is related to the semantics of its dominating prepositional phrase. Knowing the sense of a prepo- sition could help to correctly classify the semantic role of the dominating preposi- tional phrase and vice versa. In this pa- per, we propose a joint probabilistic model for word sense disambiguation of preposi- tions and semantic role labeling of prepo- sitional phrases. Our experiments on the PropBank corpus show that jointly learn- ing the word sense and the semantic role leads to an improvement over state-of-the- art individual classifier models on the two tasks. 1 Introduction Word sense disambiguation (WSD) and seman- tic role labeling (SRL) are two key components in natural language processing to find a semantic representation for a sentence. Semantic role la- beling is the task of determining the constituents of a sentence that represent semantic arguments with respect to a predicate and labeling each with a semantic role. Word sense disambiguation tries to determine the correct meaning of a word in a given context. Ambiguous words occur frequently in normal English text. One word class which is both frequent and highly ambiguous is preposition. The different senses of a preposition express different relations between the preposition complement and the rest of the sentence. Semantic roles and word senses offer two different inventories of “meaning” for prepositional phrases (PP): semantic roles distin- guish between different verb complements while word senses intend to fully capture the preposition semantics at a more fine-grained level. In this pa- per, we use the semantic roles from the PropBank corpus and the preposition senses from the Prepo- sition Project (TPP). Both corpora are explained in more detail in the following section. The re- lationship between the two inventories (PropBank semantic roles and TPP preposition senses) is not a simple one-to-one mapping, as we can see from the following examples: • She now lives with relatives [insense1 Alabama.]ARGM-LOC • The envelope arrives [insense1 the mail.]ARG4 • [Insense5 separate statements]ARGM-LOC the two sides said they want to have “further discus- sions.” In the first two examples, the sense of the preposi- tion in is annotated as sense 1 (“surrounded by or enclosed in”), following the definitions of the TPP, but the semantic roles are different. In the first example the semantic role is a locative adjunctive argument (ARGM-LOC), while in the second ex- ample it is ARG4 which denotes the “end point or destination” of the arriving action1. In the first and third example, the semantic roles are the same, but the preposition senses are different, i.e., sense 1 and sense 5 (“inclusion or involvement”). Preposition senses …

Content truncated to 500 words