Learning compositional semantics for open domain semantic parsing

This paper introduces a new approach to learning compositional semantics for open domain semantic parsing. Our approach is called Dependency-based Semantic Composition using Graphs (DeSCoG) and deviates from existing approaches in several ways. First, we remove the need of the lambda calculus by using a graph-based variant of Discourse Representation Structures to represent semantic building blocks and defining new combinatory operations for our graph structures. Second, we propose a probability model to approximate probability distributions over possible semantic compositions. And third, we use a variant of alignment algorithms from machine translation to learn a lexicon. On the Groningen Meaning Bank (a recently released, large-scale, domain-general, semantically annotated corpus; Basile et al. (2012)), where we preprocess sentences with an existing dependency parser, we achieve results significantly better than the baseline. On Geoquery we obtain performance comparable to semantic parsers that were developed specifically for that domain.