Structural disambiguation using statistical information and the measurement of confidence in attachment

In this paper, we propose a statistical approach to structural disambiguation of verbal phrase attachment. Instead of handcoded semantic knowledge that raises knowledge acquisition bottleneck problem, collocations are proposed to be used for structural disambiguation. Since some types of structural ambiguity cannot be resolved using simple collocations, they are augmented with information-theoretic concept of mutual information. The augmented collocations can be automatically acquired from a corpus. From the concept of mutual information, a new concept of confidence measure is derived that can be used for indicating how reliable the attachment is.<>