Thesauruses for Prepositional Phrase Attachment

Thesauruses for Prepositional Phrase Attachment Mark McLauchlan Department of Informatics University of Sussex Brighton, BN1 9RH mrm21@sussex.ac.uk Abstract Probabilistic models have been effective in re- solving prepositional phrase attachment am- biguity, but sparse data remains a significant problem. We propose a solution based on similarity-based smoothing, where the proba- bility of new PPs is estimated with informa- tion from similar examples generated using a thesaurus. Three thesauruses are compared on this task: two existing generic thesauruses and a new specialist PP thesaurus tailored for this problem. We also compare three smoothing techniques for prepositional phrases. We find that the similarity scores provided by the the- saurus tend to weight distant neighbours too highly, and describe a better score based on the rank of a word in the list of similar words. Our smoothing methods are applied to an existing PP attachment model and we obtain significant improvements over the baseline. 1 Introduction Prepositional phrases are an interesting example of syn- tactic ambiguity and a challenge for automatic parsers. The ambiguity arises whenever a prepositional phrase can modify a preceding verb or noun, as in the canoni- cal example I saw the man with the telescope. In syn- tactic terms, the prepositional phrase attaches either to the noun phrase or the verb phrase. Many kinds of syn- tactic ambiguity can be resolved using structural infor- mation alone (Briscoe and Carroll, 1995; Lin, 1998a; Klein and Manning, 2003), but in this case both candidate structures are perfectly grammatical and roughly equally likely. Therefore ambiguous prepositional phrases re- quire some kind of additional context to disambiguate correctly. In some cases a small amount of lexical knowl- edge is sufficient: for example of almost always modifies the noun. Other cases, such as the telescope example, are potentially much harder since discourse or world knowl- edge might be required. Fortunately it is possible to do well at this task just by considering the lexical preferences of the words mak- ing up the PP. Lexical preferences describe the tendency for certain words to occur together or only in specific constructions. For example, saw and telescope are more likely to occur together than man and telescope, so we can infer that the correct attachment is likely to be ver- bal. The most useful lexical preferences are captured by the quadruple (v, n1, p, n2) where v is the verb, n1 is the head of the direct object, p is the preposition and n2 is the head of the prepositional phrase. A benchmark dataset of 27,937 such quadruples was extracted from the Wall Street Journal corpus by Ratnaparkhi et al. (1994) and has been the basis of many subsequent studies comparing machine learning algorithms and lexical resources. This paper examines the effect of particular smoothing algo- rithms on the performance of an existing statistical PP model. A major problem faced by any statistical attachment al- gorithm is sparse data, which occurs when plausible PPs are not well-represented in the training data. For exam- ple, if the observed frequency …

Content truncated to 500 words