Controlled-Dialogue-Generation

This repository contains the data and code for the paper "SideControl: Controlled Open-domain Dialogue Generation via Additive Side Networks" (EMNLP2021-Findings).

Environment

Under the server environment of python=3.6 and CUDA 11.1, install the following packages:

pip install -r requirements.txt

Training

Training the DialoGPT-SideNet on DailyDialog full training set includes two steps. First, pretrain a DA classifier:

python gpt2_da_sidenet.py -d dailydialog_raw -t train -f base --pretrain_clf

Second, train a SideNet (remember to replace the timestamp and ckpt with your own model checkpoint):

python gpt2_da_sidenet.py -d dailydialog_raw -t ft -f sidenet --timestamp 2021-05-11-07-08-10 --ckpt 2500

Training the DialoGPT-SideNet on ConvAI2 full training set includes one step:

python gpt2_kb_sidenet.py -d convai2_raw -t train -f sidenet

Decoding

Get the DialoGPT-SideNet predictions on DailyDialog full testing set (remember to replace the timestamp and ckpt with your own model checkpoint):

python gpt2_da_sidenet.py -d dailydialog_raw -t eval -f sidenet --timestamp 2021-05-11-07-08-10 --ckpt 2500

Get the DialoGPT-SideNet predictions on ConvAI2 full testing set (remember to replace the timestamp and ckpt with your own model checkpoint):

python gpt2_kb_sidenet.py -d convai2_raw -t eval -f sidenet --timestamp 2021-04-26-10-21-06 --ckpt 47839

Evaluation

Compute text quality metrics for DialoGPT-SideNet predictions:

python evaluation.py \
--mode sent \
--reference_file {args.model_name}_{args.dataset}_{args.flag}_{args.timestamp}/refs.json \
--output_file {args.model_name}_{args.dataset}_{args.flag}_{args.timestamp}/outs.json

Compute text controllability metrics (knowledge document control) for DialoGPT-SideNet predictions:

python evaluation.py \
--mode kb \
--reference_file {args.model_name}_{args.dataset}_{args.flag}_{args.timestamp}/refs.json \
--output_file {args.model_name}_{args.dataset}_{args.flag}_{args.timestamp}/outs.json

Compute text controllability metrics (semantic label control) for DialoGPT-SideNet predictions. First, train an independent DA classifier:

python bert_da_eval.py -d dailydialog_dis -t train -f clf

Second, compute the accuracy predicted by the independent DA classifier (remember to replace the timestamp, ckpt and output_file accordingly):

python bert_da_eval.py -d dailydialog_dis -t pred -f clf \
--timestamp 2021-04-11-05-57-18 --ckpt 10000 \
--output_file {args.model_name}_{args.dataset}_{args.flag}_{args.timestamp}/outs.json

Citation

Please cite our work if you are interested.

@inproceedings{du-ji-2021-sidecontrol-controlled,
    title = "{S}ide{C}ontrol: Controlled Open-domain Dialogue Generation via Additive Side Networks",
    author = "Du, Wanyu  and
      Ji, Yangfeng",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2021",
    month = nov,
    year = "2021",
    address = "Punta Cana, Dominican Republic",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.findings-emnlp.188",
    pages = "2175--2194",
}