Measuring and Improving Semantic Diversity of Dialogue Generation (ENMLP 2022 Findings)

Paper link

• https://arxiv.org/abs/2210.05725

Overview

• codes/evalutate_diversify.py - Measuring semantic diversity of dialogue generation.
• codes/train_dress.py, train_diversify_base.py - Training the model to improve the semantic diversity of dialogue generation.
• codes/dm_generator.py - A generator agent for DRESS.

Installation

• First, this code is implemented based on ParlAI and Huggingface Transformers. You need to install ParlAI and Huggingface Transformers as described in the README on those links.

• After installing ParlAI on your local, then move the codes as follows:

  • codes/evalutate_diversify.py -> parlai/scripts/evalutate_diversify.py
  • codes/train_dress.py -> parlai/scripts/train_dress.py
  • codes/train_diversify_base.py -> parlai/scripts/train_diversify_base.py
  • codes/dm_generator.py -> parlai/agents/transformer/dm_generator.py
  • codes/utils/ -> parlai/utils

• Install mauve package in ./mauve:

  pip install -e mauve

Training & Evaluation

Training

# Blender 90M
bash scripts/train_balancing.sh [init_model] 32 7e-6 dailydialog

# Bart-large
bash scripts/train_balancing_bart.sh [init_model] 16 7e-6 dailydialog

Evaluation

bash scripts/eval_dailydialog.sh [model_filename] [report_filename] 32

Citation

Please cite the following if you make use of this codebase.

@article{han2022measuring,
  title={Measuring and Improving Semantic Diversity of Dialogue Generation},
  author={Han, Seungju and Kim, Beomsu and Chang, Buru},
  journal={arXiv preprint arXiv:2210.05725},
  year={2022}
}