Combining Unsupervised and Supervised Methods for PP Attachment Disambiguation

Combining unsupervised and supervised methods for PP attachment disambiguation Martin Volk University of Zurich Sch¨onberggasse 9 CH-8001 Zurich vlk@zhwin.ch Abstract Statistical methods for PP attachment fall into two classes according to the training material used: first, unsupervised methods trained on raw text corpora and second, supervised meth- ods trained on manually disambiguated exam- ples. Usually supervised methods win over un- supervised methods with regard to attachment accuracy. But what if only small sets of manu- ally disambiguated material are available? We show that in this case it is advantageous to in- tertwine unsupervised and supervised methods into one disambiguation algorithm that outper- forms both methods used alone.1 1 Introduction Recently, numerous statistical methods for prepositional phrase (PP) attachment disam- biguation have been proposed. They can broadly be divided into unsupervised and su- pervised methods. In the unsupervised methods the attachment decision is based on information derived from large corpora of raw text. The text may be automatically processed (e.g. by shallow parsing) but not manually disambiguated. The most prominent unsupervised methods are the Lexical Association score by Hindle and Rooth (1993) and the cooccurrence values by Ratna- parkhi (1998). They resulted in up to 82% cor- rect attachments for a set of around 3000 test cases from the Penn treebank. Pantel and Lin (2000) increased the training corpus, added a collocation database and a thesaurus which im- proved the accuracy to 84%. In contrast, the supervised methods are based on information that the program learns from manually disambiguated cases. These cases 1This research was supported by the Swiss National Science Foundation under grant 12-54106.98. are usually extracted from a treebank. Su- pervised methods are as varied as the Back- off approach by Collins and Brooks (1995) and the Transformation-based approach by Brill and Resnik (1994). Back-off scored 84% correct attachments and outperformed the Transformation-based approach (80%). Even better results were reported by Stetina and Na- gao (1997) who used the WordNet thesaurus with a supervised learner and achieved 88% ac- curacy. All these accuracy figures were reported for English. We have evaluated both unsupervised and supervised methods for PP attachment dis- ambiguation in German. This work was con- strained by the availability of only a small Ger- man treebank (10,000 sentences). Under this constraint we found that an intertwined combi- nation of using information from unsupervised and supervised learning leads to the best re- sults. We believe that our results are relevant to many languages for which only small treebanks are available. 2 Our training resources We used the NEGRA treebank (Skut et al., 1998) with 10,000 sentences from German news- papers and extracted 4-tuples (V, N1, P, N2) whenever a PP with the preposition P and the core noun N2 immediately followed a noun N1 in a clause headed by the verb V . For example, the sentence In Deutschland ist das Ger¨at ¨uber die Bad Homburger Ergos zu beziehen. [In Germany the appliance may be ordered from Er- gos based in Bad Homburg.] leads to the 4-tuple (beziehen, …

Content truncated to 500 words