ConvADR-QA: Open-Domain Conversational Question Answering with Historical Answers

• AACL-IJCNLP 2022 Findings Paper
• Our code is based on ConvDR and ORConvQA.

Framework

Installation

• It is recommended to create a conda environment for the project with conda create -n conadr_qa python=3.8
• Install dependencies

# Install torch (please check your CUDA version)
pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu102

git clone https://github.com/MiuLab/ConvADR-QA.git
cd ConvADR-QA 
pip install -r requirements.txt

Data

• Download OR-QuAC data

bash scripts/download.sh

• Preprocessing

bash scripts/preprocessing.sh

• Document Embedding Geneation

bash scripts/gen_embed.sh

Note that this step could take a lot of memory and time.

Model Training

Training

• To use ranking loss, we first need to find negative documents using manual queries:

bash scripts/run_train.sh find_neg

• After the retrieval finishes, we can select negative documents using the following script:

bash scripts/run_train.sh gen_rank

• Model training:

bash scripts/run_train.sh train

Inference

bash scripts/run_inference.sh

Citation

@inproceedings{fang2022open,
  title={Open-Domain Conversational Question Answering with Historical Answers},
  author={Fang, Hung-Chieh and Hung, Kuo-Han and Huang, Chen-Wei and Chen, Yun-Nung},
  booktitle={Findings of the Association for Computational Linguistics: AACL-IJCNLP 2022},
  pages={319--326},
  year={2022}
}