Semantic Mapping for Lexical Sparseness Reduction in Parsing

Bilexical information is known to be helpful in parse disambiguation, but the benefit is limited because of lexical sparseness. An approach using word classes can reduce sparseness and potentially leads to more accurate parsing. Firstly, we describe a method identifying the dependency types of the Alpino parser for Dutch to which we would like to apply generalization. These are the types which are most likely to reduce the sparseness and positively affect parsing at the same time. Secondly, we provide preliminary results for enhancement of dependency types with semantic classes derived from a WordNet-like inventory for Dutch. Classes of varying degrees of generality are applied to three dependency types: nominal conjunction, modification of adjective and modification of noun. We observe improvements in some concrete cases, whereas the overall parsing accuracy either remains unchanged or decreases. We identify drawbacks of human-built sense inventories, which provides motivation for a distributional semantic approach.