This is a Python implementation of the work Learning Interpretable Latent Dialogue Actions With Less Supervision. For questions or comments, contact [email protected].
We implement several readers for different datasets. Namely it is:
You can use Makefile to download & prepare the data:
make DATA_DIR=data_dir_path [camrest|smd|multiwoz]
If you cannot/do not want to use the Makefile, please refer to the script to prepare the data manually.
The codebase is written in Python 3.6 and requires the dependencies specified at requirements.txt file.
To install, simply run in your environment:
pip install -r requirements.txt
To train the model, first prepare the configuration.
Sample configuration is located at config/vrnn.yaml.
After the configuration is prepared, run:
python -m VRNN-generation.run VRNN-generation/config/vrnn.yaml --output VRNN-generation/output/$DATA
where $DATA=[camrest|multiwoz|smd]
To evaluate various metrics, you can use the prepares evaluation script as follows:
python -m evaluation.run --work_dir output_model_dir --fn model_predictions.txt --metrics $METRIC_MODE
where $METRIC_MODE can be one or more values from the following list:
- bleu Evaluate the BLEU score
- z_semantics Train and evaluate a DT classifier that predicts system actions from latent variables.
- z_info Compute Mutual Information between the latent variables and ground truth actions.
- ppl Compute perplexity
- success Compute (modified) dialogue success rate
- emr Compute Entity Match Rate
To cite this work, please use the following record:
@misc{https://doi.org/10.48550/arxiv.2209.11128,
doi = {10.48550/ARXIV.2209.11128},
url = {https://arxiv.org/abs/2209.11128},
author = {Hudeček, Vojtěch and Dušek, Ondřej},
keywords = {Computation and Language (cs.CL), FOS: Computer and information sciences, FOS: Computer and information sciences},
title = {Learning Interpretable Latent Dialogue Actions With Less Supervision},
publisher = {arXiv},
year = {2022},
copyright = {arXiv.org perpetual, non-exclusive license}
}
This work was partially supported by Charles University projects PRIMUS/19/SCI/10, GA UK No. 302120 and SVV No. 260575 and by the European Research Council (Grant agreement No. 101039303 NG-NLG). It used resources provided by the LINDAT/CLARIAH-CZ Research Infrastructure (Czech Ministry of Education, Youth and Sports project No. LM2018101).