DSI

This work is based on the paper Transforming Slot Schema Induction with Generative Dialogue State Inference, which trains a S2S model to discover new slots from unlabeled dialogue data. The S2S model is trained on the D0T dataset.

Install

Create a conda environment then pip install requirements.txt

conda create --solver=libmamba -n dsi -c rapidsai -c conda-forge -c nvidia cudf=24.02 cuml=24.02 python=3.10 cuda-version=11.8
conda activate dsi
pip install -r requirements.txt

DSI T5 model

The training/experiment code relies on loading a pickle object representing the D0T data. Therefore, a simpler usage of the DSI model is exemplified in dsi/s2s_dsi.py. The folder s2s_dsi/ contains the training and experiment code for reference.

Full SSI

After getting labels from the DSI s2s model (see data/silver_...), SBERT encoding and clustering are used to induce the final schema. Experiment code for this is in dsi, with dsi/experiment.py being the main script.