The proposed project seeks to target key open questions in language structure and use by bringing together researchers in three areas of language research, in all of which MIT is already a world leader: statistical natural language processing, formal linguistics, and psycholinguistics.
Our first proposed collaboration furthers the use of linguistic insights to support NLP modeling. We propose to develop a novel inference algorithm that induces linguistic structures satisfying constraints encoded by linguistic experts. During the learning process, the algorithm will automatically reconcile these constraints with the patterns observed in the data. This process increases coverage and augments human-specified rules and representations with probabilities, while simultaneously constraining the automatically induced structures to be consistent with linguistic knowledge. To achieve this learning behavior, we propose devising a novel variational inference procedure that concentrates exploration in areas of the parameter space that contain structures consistent with human-specified constraints.