This repository provides the implementation details for the ACL 2021 main conference paper:

BoB: BERT Over BERT for Training Persona-based Dialogue Models from Limited Personalized Data. [paper]

In this work, we carried out persona-based dialogue generation experiments under a persona-dense scenario (English PersonaChat) and a persona-sparse scenario (Chinese PersonalDialog), with the assistance of a series of auxiliary inference datasets. Here we summarize the key information of these datasets and provide the links to download these datasets if they are directly accessible.

• For Persona-Dense Experiments

  Dataset	Type	Language	Usage	Where to Download
  ConvAI2 PersonaChat	Dialogue Generation	English	Training	https://www.aclweb.org/anthology/P18-1205.pdf train_self_original_no_cands & valid_self_original_no_cands (7801 test dialogues)
  MNLI	Non-dialogue Inference	English	Training	https://cims.nyu.edu/~sbowman/multinli/multinli_1.0.zip entailment & contradiction
  DNLI	Dialogue Inference	English	Evaluation	https://www.aclweb.org/anthology/P19-1363.pdf

• For Persona-Sparse Experiments

  Dataset	Type	Language	Usage	Where to Download
  ECDT2019 PersonalDialog	Dialogue Generation	Chinese	Training	https://arxiv.org/pdf/1901.09672.pdf dialogues_train.json & test_data_random.json & test_data_biased.json
  CMNLI	Non-dialogue Inference	Chinese	Training	https://github.com/CLUEbenchmark/CLUECorpus2020/ entailment & contradiction
  KvPI	Dialogue Inference	Chinese	Evaluation	https://github.com/songhaoyu/KvPI

• Download Pre-trained BERT

  The BoB model is initialized from public BERT checkpoints:

  • English BERT: https://huggingface.co/bert-base-uncased/tree/main
  • Chinese BERT: https://huggingface.co/bert-base-chinese/tree/main

The setup.sh script contains the necessary dependencies to run this project. Simply run ./setup.sh would install these dependencies. Here we take the English PersonaChat dataset as an example to illustrate how to run the dialogue generation experiments. Generally, there are three steps, i.e., tokenization, training and inference:

• Preprocessing

   python preprocess.py --dataset_type convai2 \
   --trainset ./data/ConvAI2/train_self_original_no_cands.txt \
   --testset ./data/ConvAI2/valid_self_original_no_cands.txt \
   --nliset ./data/ConvAI2/ \
   --encoder_model_name_or_path ./pretrained_models/bert/bert-base-uncased/ \
   --max_source_length 64 \
   --max_target_length 32
  

  We have provided some data examples (dozens of lines) at the ./data directory to show the data format. preprocess.py reads different datasets and tokenizes the raw data into a series of vocab IDs to facilitate model training. The --dataset_type could be either convai2 (for English PersonaChat) or ecdt2019 (for Chinese PersonalDialog). Finally, the tokenized data will be saved as a series of JSON files.

• Model Training

   CUDA_VISIBLE_DEVICES=0 python bertoverbert.py --do_train \
   --encoder_model ./pretrained_models/bert/bert-base-uncased/ \
   --decoder_model ./pretrained_models/bert/bert-base-uncased/ \
   --decoder2_model ./pretrained_models/bert/bert-base-uncased/ \
   --save_model_path checkpoints/ConvAI2/bertoverbert --dataset_type convai2 \
   --dumped_token ./data/ConvAI2/convai2_tokenized/ \
   --learning_rate 7e-6 \
   --batch_size 32
  

  Here we initialize encoder and both decoders from the same downloaded BERT checkpoint. And more parameter settings could be found at bertoverbert.py.

• Evaluations

   CUDA_VISIBLE_DEVICES=0 python bertoverbert.py --dumped_token ./data/ConvAI2/convai2_tokenized/ \
   --dataset_type convai2 \
   --encoder_model ./pretrained_models/bert/bert-base-uncased/  \
   --do_evaluation --do_predict \
   --eval_epoch 7
  

  Empirically, in the PersonaChat experiment with default hyperparameter settings, the best-performing checkpoint should be found between epoch 5 and epoch 9. If the training procedure goes fine, there should be some results like:

   Perplexity on test set is 21.037 and 7.813.
  

  where 21.037 is the ppl from the first decoder and 7.813 is the final ppl from the second decoder. And the generated results is redirected to test_result.tsv, here is a generated example from the above checkpoint:

   persona:i'm terrified of scorpions. i am employed by the us postal service. i've a german shepherd named barnaby. my father drove a car for nascar.
   query:sorry to hear that. my dad is an army soldier.
   gold:i thank him for his service.
   response_from_d1:that's cool. i'm a train driver.
   response_from_d2:that's cool. i'm a bit of a canadian who works for america.  
  

  where d1 and d2 are the two BERT decoders, respectively.

• Computing Infrastructure:

  • The released codes were tested on NVIDIA Tesla V100 32G and NVIDIA PCIe A100 40G GPUs. Notice that with a batch_size=32, the BoB model will need at least 20Gb GPU resources for training.

• Build upon 🤗 Transformers.

• Bibtex:

    @inproceedings{song-etal-2021-bob,
        title = "{B}o{B}: {BERT} Over {BERT} for Training Persona-based Dialogue Models from Limited Personalized Data",
        author = "Song, Haoyu  and
          Wang, Yan  and
          Zhang, Kaiyan  and
          Zhang, Wei-Nan  and
          Liu, Ting",
        booktitle = "Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)",
        month = aug,
        year = "2021",
        address = "Online",
        publisher = "Association for Computational Linguistics",
        url = "https://aclanthology.org/2021.acl-long.14",
        doi = "10.18653/v1/2021.acl-long.14",
        pages = "167--177",
    }
    

• Email: hysong@ir.hit.edu.cn.