Symmetry-Based Semantic Parsing

Semantic parsing maps sentences to formal meaning representations, enabling question answering, natural language interfaces, and many other applications. However, there is no agreement on what the meaning representation should be, and constructing a sufficiently large corpus of sentence-meaning pairs for learning is extremely challenging. In this paper, we argue that both of these problems can be avoided if we adopt a new notion of semantics. For this, we take advantage of symmetry group theory, a highly developed area of mathematics concerned with transformations of a structure that preserve its key properties. We define a symmetry of a sentence as a syntactic transformation that preserves its meaning. Semantically parsing a sentence then consists of inferring its most probable orbit under the language’s symmetry group, i.e., the set of sentences that it can be transformed into by symmetries in the group. The orbit is an implicit representation of a sentence’s meaning that suffices for most applications. Learning a semantic parser consists of discovering likely symmetries of the language (e.g., paraphrases) from a corpus of sentence pairs with the same meaning. Once discovered, symmetries can be composed in a wide variety of ways, potentially resulting in an unprecedented degree of immunity to syntactic variation.