Abstract
A recent development in natural language semantics combines a quantitative, context-dependent model of lexical meanings with a compositional account of how meanings combine into larger units. Existing models for this composition process suffer from under- and over generation. These shortcomings are particularly felt in languages with entangled dependency structures, such as Dutch. We develop a type-driven composition calculus with fine control over locally restricted and non-local forms of information flow. The composition calculus is language-independent; localisations are obtained by combining it with a language-specific lexicon. We induce a Dutch type lexicon that captures the required dependency patterns from treebank data. We perform experiments to highlight the enhanced empirical performance of the proposed model for vector-based semantic modelling. The results of the project add a key component to the digital infrastructure of Dutch, enabling deep semantic analysis of unstructured data.