CausalDialogue

This repository contains code and data used in this paper. The code part is under MIT License; and the data part, as described in data/README.md is under GNU Free Document License. We have included the license documents in this repository. If you find the code or data is useful for your research, please kindly cite the paper.

Setup

• clone this repository, you might also need to refer to Git LFS in order to download the data.
• (recommend) create a conda environment.
• install packages via conda and pip install.

$conda install graphviz
$pip install -r requirements.txt

• install required nltk package

$python
>>> import nltk
>>> nltk.download('punkt')

Usage

• Training; an example of fine-tuning a T5 model on CausalDialog using MLE

python main.py --task fethco --model_type t5 --lr 1e-5 --n_epochs 5 --train_batch_size 16 --valid_batch_size 32 --gradient_accumulation_steps 4 --max_history_len 256 --model_checkpoint exp_t5_standard --do_train

• Testing; an example of generation on test set using TopK sampling

python main.py --task fethco --model_type t5 --model_checkpoint exp_t5_standard/pytorch_model.bin --max_history_len 256 --do_test --do_generate --loss_type standard --sample_method topk --test_batch_size 128 --preds_outpath exp_t5_standard/ckpt_last_topk

• To get automatic evaluation results, we currently follow three steps.
  • first utilize visualize_gt_data.py to have the references.
  • second utilize gt_eval.py to get some reference of the ground-truths results.
  • third utilize post_eval.py to get the results of the generated file in the testing stage.
• Some notes:
  • --task fethco is an alias and required for the main.py. For now, there is no other tasks written in the program.
  • --model_type can be set to t5 or dialogpt for now.
  • --loss_type can be set to standard or exmate for now.
  • --sample_method can be set to argmax, softmax, or topk for now.