Preposition Semantic Classification via Treebank and FrameNet

Preposition Semantic Classification via PENN TREEBANK and FRAMENET Tom O’Hara Department of Computer Science New Mexico State University Las Cruces, NM 88003 tomohara@cs.nmsu.edu Janyce Wiebe Department of Computer Science University of Pittsburgh Pittsburgh, PA 15260 wiebe@cs.pitt.edu Abstract This paper reports on experiments in clas- sifying the semantic role annotations as- signed to prepositional phrases in both the PENN TREEBANK and FRAMENET. In both cases, experiments are done to see how the prepositions can be classified given the dataset’s role inventory, using standard word-sense disambiguation fea- tures. In addition to using traditional word collocations, the experiments incorporate class-based collocations in the form of WordNet hypernyms. For Treebank, the word collocations achieve slightly better performance: 78.5% versus 77.4% when separate classifiers are used per preposi- tion. When using a single classifier for all of the prepositions together, the com- bined approach yields a significant gain at 85.8% accuracy versus 81.3% for word- only collocations. For FrameNet, the combined use of both collocation types achieves better performance for the indi- vidual classifiers: 70.3% versus 68.5%. However, classification using a single classifier is not effective due to confusion among the fine-grained roles. 1 Introduction English prepositions convey important relations in text. When used as verbal adjuncts, they are the prin- ciple means of conveying semantic roles for the sup- porting entities described by the predicate. Preposi- tions are highly ambiguous. A typical collegiate dic- tionary has dozens of senses for each of the common prepositions. These senses tend to be closely related, in contrast to the other parts of speech where there might be a variety of distinct senses. Given the recent advances in word-sense disam- biguation, due in part to SENSEVAL (Edmonds and Cotton, 2001), it would seem natural to apply the same basic approach to handling the disambiguation of prepositions. Of course, it is difficult to disam- biguate prepositions at the granularity present in col- legiate dictionaries, as illustrated later. Nonetheless, in certain cases this is feasible. We provide results for disambiguating preposi- tions at two different levels of granularity. The coarse granularity is more typical of earlier work in computational linguistics, such as the role inventory proposed by Fillmore (1968), including high-level roles such as instrument and location. Recently, sys- tems have incorporated fine-grained roles, often spe- cific to particular domains. For example, in the Cyc KB there are close to 200 different types of seman- tic roles. These range from high-level roles (e.g., beneficiaries) through medium-level roles (e.g., ex- changes) to highly specialized roles (e.g., catalyst).1 Preposition classification using two different se- mantic role inventories are investigated in this pa- per, taking advantage of large annotated corpora. After providing background to the work in Sec- tion 2, experiments over the semantic role anno- tations are discussed in Section 3. The results over TREEBANK (Marcus et al., 1994) are covered first. Treebank include about a dozen high-level roles similar to Fillmore’s. Next, experiments us- ing the finer-grained semantic role annotations in FRAMENET version 0.75 (Fillmore et al., 2001) are 1Part of …

Content truncated to 500 words