Skip to content

CONVERSER: Few-Shot Conversational Dense Retrieval with Synthetic Data Generation, SIGDIAL 2023

Notifications You must be signed in to change notification settings

MiuLab/CONVERSER

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 

Repository files navigation

CONVERSER: Few-Shot Conversational Dense Retrieval with Synthetic Data Generation

SIGDIAL 2023 paper image

Reference

Please cite the following paper

@inproceedings{huang2023converser,
    title = "{CONVERSER}: Few-shot Conversational Dense Retrieval with Synthetic Data Generation",
    author = "Huang, Chao-Wei and Hsu, Chen-Yu and Hsu, Tsu-Yuan and Li, Chen-An and Chen, Yun-Nung",
    booktitle = "Proceedings of the 24th Annual Meeting of the Special Interest Group on Discourse and Dialogue",
    month = sep,
    year = "2023",
    address = "Prague, Czechia",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.sigdial-1.34",
    doi = "10.18653/v1/2023.sigdial-1.34",
    pages = "381--387"
}

Requirements

  • Python >= 3.6
  • Transformers
  • torch

Datasets

Our generated dataset can be found in the google drive

How to run

Pretrained LLM

We used LLaMA-13B in our experiments. Please apply for access here. You can also try other open-source LLMs such as LLaMA-2 and Falcon. Note that our method doesn't require instruction-tuned LLMs, so you can use any pretrained LLM.

Corpus

In order to run dialogue generation, you'll need a collection of passages. In our experiments, we used the passage collection from OR-QuAC. You can process the released data with the ConvDR repo.

Run the generation script

  • Modify the paths to LLAMA_CHECKPOINT_DIR and COLLECTION_JSONL in generate_dialog.py to your local paths.
  • Simple run
    python3 generate_dialog.py
  • You can also find our generated datasets here

Training a DPR model

Please refer to the original DPR repo or the more resource-light implementation GC-DPR for training a DPR model given the generated dataset. With GC-DPR, you should be able to train a DPR model with only 1 GPU. Below is a reference command we used with GC-DPR to train the model:

CUDA_VISIBLE_DEVICES=0 python3 train_dense_encoder.py \
    --max_grad_norm 2.0 \
    --encoder_model_type hf_bert \
    --pretrained_model_cfg bert-base-uncased \
    --seed 12345 \
    --sequence_length 384 \
    --warmup_steps 1237 \
    --batch_size 64 \
    --dev_batch_size 16 \
    --do_lower_case \
    --train_file ${GENERATED_DATASET} \
    --dev_file ../ConvDR/datasets/or-quac/dev_dpr.json \
    --output_dir ${MODEL_DIR} \
    --learning_rate 2e-05 \
    --num_train_epochs 30 \
    --val_av_rank_start_epoch 0 \
    --fp16 \
    --grad_cache \
    --q_chunk_size 8 \
    --ctx_chunk_size 8 \
    --global_loss_buf_sz 2097152 \
    --val_av_rank_max_qs 10000

About

CONVERSER: Few-Shot Conversational Dense Retrieval with Synthetic Data Generation, SIGDIAL 2023

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages