This repository provides the Pytorch source code for Transformer-based span-selection for goal-oriented information-seeking dialogue systems. In particular, the goal is to teach a dialogue system to identify the most relevant knowledge in the associated document for generating agent responses in natural language. Details are described in doc2dial: A Goal-Oriented Document-Grounded Dialogue Dataset, in EMNLP 2020.

Built with transformers from HuggingFace🤗. (Thanks HuggingFace!)

The Doc2Dial dataset contains goal-oriented conversations between an end user and an assistive agent. Each turn in a conversation is annotated with a dialogue scene, which includes speaker role, dialogue act, and grounding in a document. Detailed description can be found at the dataset webpage here.

Please use the bibtex entry below to cite our paper if you use the dataset or the baseline code. Thank you!

@inproceedings{feng-etal-2020-doc2dial,
    title = "doc2dial: A Goal-Oriented Document-Grounded Dialogue Dataset",
    author = "Feng, Song  and Wan, Hui  and Gunasekara, Chulaka  and Patel, Siva  and Joshi, Sachindra  and Lastras, Luis",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
    month = nov,
    year = "2020",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.emnlp-main.652",
}

• Python (3.7 or higher)
• Pytorch (1.1.0 or higher)
• HuggingFace / Transformer (3.3.0)
• Numpy
• tensorboard
• tqdm

These can be installed using pip by running :

pip install -r requirements.txt

Example command to train a model on Doc2Dial training set (with distributed training on 2 V100 GPUs):

python -m torch.distributed.launch --nproc_per_node=2 knowledge_identification/run_doc2dial.py 
    --model_type bert 
    --model_name_or_path bert-base-uncased
    --do_train 
    --evaluate_during_training
    --logging_steps 500
    --save_steps 500
    --do_lower_case 
    --data_dir $datadir
    --doc_file doc2dial_doc_data.json 
    --train_dial_file doc2dial_dial_data_train.json 
    --eval_dial_file doc2dial_dial_data_dev.json 
    --learning_rate 3e-5 
    --num_train_epochs 3
    --max_seq_length 512 
    --doc_stride 128
    --output_dir $output_dir
    --overwrite_output_dir 
    --per_gpu_eval_batch_size=27   
    --per_gpu_train_batch_size=27   
    --gradient_accumulation_steps=1 
    --warmup_steps=500
    --weight_decay=0.01
    --fp16
    --get_utterances all
    --max_query_length 256

For task to identify knowledge grounding for agent turns instead of user turns, add --predict_agent.

Example command to decode Doc2Dial test set with a model trained with the above setting:

python  $codedir/knowledge_identification/run_doc2dial.py 
    --model_type bert 
    --model_name_or_path $model_dir 
    --do_eval 
    --do_lower_case 
    --data_dir $datadir
    --doc_file doc2dial_doc_data.json 
    --eval_dial_file doc2dial_dial_data_test.json 
    --max_seq_length 512 
    --doc_stride 128
    --output_dir $output_dir
    --overwrite_output_dir 
    --per_gpu_eval_batch_size=27 
    --fp16
    --get_utterances all
    --max_query_length 256

For task to identify knowledge grounding for agent turns instead of user turns, add --predict_agent.