CLIN 2005 Abstracts
  • Prepositional Phrase Attachment for Dutch: New attention for an old task
    Hans van Halteren (Language and Speech, Radboud University Nijmegen)
    Peter Arno Coppen (Language and Speech, Radboud University Nijmegen
    Bram Elffers (Language and Speech, Radboud University Nijmegen)
    Dusan Bavcar (Language and Speech, Radboud University Nijmegen)
    Micha Hulsbosch (Language and Speech, Radboud University Nijmegen)
    The attachment of prepositional phrases to the correct node in a parse tree is one of the traditional NLP tasks on which various methods and machine learning features can be tested. Although lately there seems to be less interest in this, at least as a separate task, we do not feel it can already be put aside as being solved. Also, in our opinion, the task should be viewed and handled differently than in its traditional form.

    In most PP attachment publications, the task consists of a choice between a noun and a verb attachment for the PP, and the choice has to be based on the lexical form of the heads of the potential attachments, the preposition and the head of the prepositional complement. The evaluation is in terms of percentage of correct choices for exactly these cases. We would rather like to investigate the attachment of all prepositions and also all potential attachment points, and measure quality in terms of precision and recall. As a benchmark for Dutch we can use (parts of) the Alpino Treebank. As for the information which can be used to base attachment decisions on, we also allow access to the words' positions within the sentence, co-occurrence statistics in a larger corpus (the Twente News Corpus) and aspects of the embedding syntactic structure as provided by the Amazon parser.

    In the paper, we present the full experimental setup, examine the relative values of the various information types, and compare the results when using the different evaluation criteria.