Understanding and Improving the Exemplar-based Generation for Open-domain Conversation

Code for the paper: "Understanding and Improving the Exemplar-based Generation for Open-domain Conversation", which is presented in ACL 2022, 4th Workshop on NLP for ConvAI as an oral paper. Our implementation is based on ParlAI, so before you start we recommend you to read the README of ParlAI and prepare for the experimental setting. You can see our implementations in parlai/agents/transformer/corge.py and parlai/agents/transformer/exemplar_based_generator.py.

Preliminaries

You should prepare the pre-trained generator and retriever model to train your exemplar-based generator. For the generator, we use zoo:blender/blender_90M/model from ParlAI, however it is okay to use the model which is pre-trained on the Pushshift. For the retriever, we fine-tuned the zoo:pretrained_transformers/bi_model_huge_reddit/model from ParlAI using BST+ dataset. We assume your generator and retriever is stored in GENERATOR_MODEL_PATH and RETRIEVER_MODEL_PATH, respectively.

Training Exemplar-based Generator

export RETRIEVER_MODEL_PATH=[];
export GENERATOR_MODEL_PATH=[];
export FIXED_CANDIDATE_PATH=[];

./scripts/train_model.sh exp 5 32 7e-6;

Evaluating Exemplar-based Generator

# You don't need to load retriever and generator after you trained your exemplar-based generator.
export FIXED_CANDIDATE_PATH=[];

./scripts/eval_model.sh [MODEL_PATH];

We also provided the automatic evaulation scripts in jupyter notebook.

Get Inference Results from the Exemplar-based Generator

# You don't need to load retriever and generator after you trained your exemplar-based generator.
export FIXED_CANDIDATE_PATH=[];

./scripts/inference_model.sh [MODEL_PATH];

Citation

If you find our paper or this project helps your research, please kindly consider citing our paper in your publications.

@article{han2021understanding,
  title={Understanding and Improving the Exemplar-based Generation for Open-domain Conversation},
  author={Han, Seungju and Kim, Beomsu and Seo, Seokjun and Erdenee, Enkhbayar and Chang, Buru},
  journal={arXiv preprint arXiv:2112.06723},
  year={2021}
}