Noun compounds are a frequently occurring yet highly ambiguous construction in natural language; their interpretation relies on extra-syntactic information. Several statistical methods for compound disambiguation have been reported in the literature; however, a striking feature of all these approaches is that disambiguation relies on statistics derived from unambiguous compounds in training, meaning they are prone to the problem of sparse data. Other researchers have overcome this difficulty somewhat by using manually crafted knowledge resources to collect statistics on "concepts" rather than noun tokens, but have sacrificed domain-independence by doing so. We report here on work investigating the application of Latent Semantic Indexing [4], an Information Retrieval technique, to the task of noun compound disambiguation. We achieved an accuracy of 84%, indicating the potential of applying vector-based distributional information measures to syntactic disambiguation.
展开▼
