Cross-lingual semantic parsing with categorial grammars

Humans communicate using natural language. We need to make sure that computers can understand us so that they can act on our spoken commands or independently gain new insights from knowledge that is written down as text. A “semantic parser” is a program that translates natural-language sentences into computer commands or logical formulas–something a computer can work with. Despite much recent progress on semantic parsing, most research focuses on English, and semantic parsers for other languages cannot keep up with the developments. My thesis aims to help close this gap. It investigates “cross-lingual learning” methods by which a computer can automatically adapt a semantic parser to another language, such as Dutch. The computer learns by looking at example sentences and their translations, e.g., “She likes to read books”/”Ze leest graag boeken”. Even with many such examples, learning which word means what and how word meanings combine into sentence meanings is a challenge, because translations are rarely word-for-word. They exhibit grammatical differences and non-literalities. My thesis presents a method for tackling these challenges based on the grammar formalism Combinatory Categorial Grammar. It shows that this is a suitable formalism for this purpose, that many structural differences between sentences and their translations can be dealt with in this framework, and that a (rudimentary) semantic parser for Dutch can be learned cross-lingually based on one for English. I also investigate methods for building large corpora of texts annotated with logical formulas to further study and improve semantic parsers.