JointBERT

Why we included this project

JointBERT is a compact, readable implementation of the classic paper that combines intent classification and slot filling in one BERT model, so a single pass predicts both what the user wants and the parts of their utterance that carry it. That joint setup is the main thing to take away: instead of training two separate classifiers and stitching them together, you get one model handling both tasks at once. The repo reproduces the ATIS and Snips benchmarks, supports several transformer backbones, and has an optional CRF layer on the slot head for enforcing valid label sequences. It is a research-style codebase pinned to older library versions, so treat it as a template to study and port rather than something to drop straight into production.