Skip to content
This repository has been archived by the owner on Jul 22, 2024. It is now read-only.
/ multidoc2dial Public archive

MultiDoc2Dial: Modeling Dialogues Grounded in Multiple Documents

License

Notifications You must be signed in to change notification settings

IBM/multidoc2dial

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

27 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

MultiDoc2Dial: Modeling Dialogues Grounded in Multiple Documents

This repository provides data and code for the corresponding paper "MultiDoc2Dial: Modeling Dialogues Grounded in Multiple Documents" (EMNLP 2021) by Song Feng , Siva Sankalp Patel, Wan Hui and Sachindra Joshi. Please cite the paper and star the repository if you find the paper, data and code useful for your work.

@inproceedings{feng2021multidoc2dial,
    title={MultiDoc2Dial: Modeling Dialogues Grounded in Multiple Documents},
    author={Feng, Song and Patel, Siva Sankalp and Wan, Hui and Joshi, Sachindra},
    booktitle={EMNLP},
    year={2021}
}

Installation

Please refer to conda_env.yml for creating a virtual environment.

conda env create -f conda_env.yml

Our scripts require to set the following environment variables,

  • HF_HOME- for caching downloads from Huggingface locally.
  • CHECKPOINTS for saving the checkpoints.

Data

Please run the commands to download data. It will download the document and dialogue data into folder data/multidoc2dial.

cd scripts
./run_download.sh

Document preprocessing

To segment the document into passages, please refer to

run_data_preprocessing.sh

Data preprocessing for fine-tuning DPR

If you are finetuning DPR on MultiDoc2Dial, please refer to run_data_preprocessing_dpr.sh create positive and negative examples in the format of DPR.

Run Baselines

Finetuning DPR

To finetune DPR, we use Facebook DPR (March 2021 release) with an effective batch size 128. You can finetune DPR on MultiDoc2Dial data yourself ; or use our finetuned version.

If you would like to finetune DPR yourself, please refer to Facebook DPR for detailed instructions.

Or

If you would like to use our finetuned DPR encoders, please use the the following paths as the model path to ctx or question encoder (for instance, run_converter_modelcard.sh),

  • sivasankalpp/dpr-multidoc2dial-token-question-encoder for fine-tuned DPR question encoder based on token-segmented document passages (link)
  • sivasankalpp/dpr-multidoc2dial-token-ctx-encoder for fine-tuned DPR ctx encoder based on token-segmented document passages (link)
  • sivasankalpp/dpr-multidoc2dial-structure-question-encoder fine-tuned DPR question encoder based on structure-segmented document passages (link)
  • sivasankalpp/dpr-multidoc2dial-structure-ctx-encoder for fine-tuned DPR ctx encoder based on structure-segmented document passages (link)

Using finetuned DPR encoders in RAG

If you obtain your own finetuned DPR checkpoints,

  1. Download the following files from RAG model cards to "../data" folder
  1. Convert your fine-tuned DPR checkpoint and add it to RAG model. Please refer to run_converter.sh.

OR

If you use our finetuned DPR encoders, please refer to run_converter_modelcard.sh.

Finetuning RAG

Our implementation is based on Huggingface RAG. Please refer to their README for more detailed explanations on document retrieval and finetuning RAG.

To create FAISS index, please refer to

run_kb_index.sh

To finetune RAG on MultiDoc2Dial data, please refer to

run_finetune_rag.sh

Evaluations

To evaluate the retrieval results (recall@n for passage and document level), please refer to

run_eval_rag_re.sh

To evaluate the generation results, please refer to

run_eval_rag_e2e.sh

Results

The evaluation results on the validation set of agent response generation task Please refer to the scripts for corresponding hyperparameters.

Model F1 EM BLEU r@1 r@5 r@10
D-token-nq 30.9 2.8 15.7 25.8 48.2 57.7
D-struct-nq 31.5 3.2 16.6 27.4 51.1 60.2
D-token-ft 33.2 3.4 18.8 35.2 63.4 72.9
D-struct-ft 33.7 3.5 19.5 37.5 67.0 75.8

Leaderboard

Please check out our leaderboard and Shared Task.

Acknowledgement

Our code is based on Huggingface Transformers. Our dataset is based on Doc2Dial. We thank the authors for sharing their great work.

About

MultiDoc2Dial: Modeling Dialogues Grounded in Multiple Documents

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 4

  •  
  •  
  •  
  •