Learning meaning from structure: neural semantic parsing with minimalist grammars


Automatically finding the meaning of a sentence (semantic parsing) is a very difficult task with critical applications in artificial intelligence. Despite this, little work has been done applying the insights of theoretical linguistics to the task. My proposal is to combine state of the art machine learning techniques with contemporary minimalist syntax to design a truly compositional and linguistically-motivated semantic parser -- one that uses the syntactic structure to inform the semantic derivation, and vice versa. In particular, I will build the first neural syntactic parsers for Minimalist Grammars, and use a deep neural network (a Tree LSTM) to recursively transform a syntactic derivation given by the parser into a semantic derivation. I will then use reinforcement learning to simultaneously train the syntactic parser to give more semantically meaningful parses, and train the Tree LSTM to enact better transformations, yielding more accurate semantic representations.
Semantic corpora exist for a variety of semantic representations (from graphs to DRSes to SQL queries); I will adapt the parser to multiple semantic representations.


Project number


Main applicant

Dr. M. Fowlie

Affiliated with

Universiteit Utrecht


01/09/2019 to 31/08/2022