Resolving PP-attachment ambiguity in French with distributional methods

The present work investigates the role of two types of distributional models in resolving the prepositional-phrase attachment ambiguity: a co-occurrence-based model and a distributional-semantic model. We show that both the attachment disambiguation and the syntactic parsing can positively benefit from distributional-semantic modeling. Our approach is original in that it uses semantic similarity as a confidence measure in choosing the correct attachment site. Complete representations of the prepositional phrase and the competing attachment sites are built from a large corpus and then used in similarity comparison, either independently or by means of vector composition. Our word-space model outperforms a chance selection of attachment sites in an isolated evaluation, but is also integrated into parsing by the process of pre-annotation of relevant dependencies, which are kept intact by the parser during parsing time. In this configuration, our model can select the correct attachment site more often than the parser, and the incorporation into parsing yields a small, but noticeable improvement of the overall parsing accuracy.