Minimal structure modeling
Existing work in probabilistic language modeling can be mostly divided into two categories:
- 2018 - 2022
- Wessel Kraaij
(i) Purely sequential, string-level approaches ensure fluency at the local level without notion of grammaticality and seek improvements in the use of massive training corpora. (ii) Fully structural, tree-based approaches model text as the realization of latent tree structures that encode complex grammatical dependencies. This project explores a third way where only structural relations required to produce grammatical sentences in a specific language and task are modeled.