Home

Awesome

HINT

A generation model equipped with HIgh-level representations for loNg Text generation described in the paper Long Text Generation by Modeling Sentence-Level and Discourse-Level Coherence (ACL 2021 Long Paper).

Prerequisites

The code is written in TensorFlow library. To use the program the following prerequisites need to be installed.

Computing infrastructure

We train HINT based on the platform:

Quick Start

1. Constructing Training Examples

The full data can be downloaded from THUcloud or GoogleDrive. The structure for the directory data is as follows

├── Data
   └── `pro_data.py`             # the code to create training examples
   └── `preprocess.sh`   # the script to create training examples
   └── `ini_data`		# the directory for the inital data
       ├── `roc`        # ROCStories
              └── `train.txt`        # the full texts including inputs and outputs (sentences separated by [SEP])
              └── `train.source`    # only inputs
              └── `train.target`       # only outputs
              └── ...
       ├── `wp`        # WritingPrompts
              └── ...
       ├── `bc`      # BookCorpus
              └── ...

   └── `data`		# training examples
       ├── `roc`        # ROCStories
              └── `train.source`    # only inputs
              └── `train.target`       # only outputs
              └── `train_order.target`       # file for recording sentence orders
              └── `train_sbertscore.target`       # file for recording the computed sbert score between sentences              
              └── ...

2. Post-Training on BookCorpus

Execute the following command (or run bash ./finetune.sh directly) to post-train BART on BookCorpus:

cd ./model
env CUDA_VISIBLE_DEVICES=0 CUDA_LAUNCH_BLOCKING=1 python3.7 -u finetune.py \
    --data_dir ../Data/data/bc \
    --output_dir=./bart_post_bc \
    --save_top_k 80 \
    --train_batch_size=10 \
    --eval_batch_size=10 \
    --num_train_epochs 10 \
    --model_name_or_path ./bart_model \
    --learning_rate=3e-5 \
    --fp16 \
    --gpus 1 \
    --do_train \
    --n_val 1000 \
    --val_check_interval 0.1 \
    --overwrite_output_dir \
    --sbert \
    --reorder \

The initial checkpoint of BART can be downloaded from BART. We use the base version of BART. We train the model for about 0.1M steps. The training process will task about 1~2 days. The post-trained model can be downloaded from THUcloud or GoogleDrive.

3. Fine-tuning on ROCStories/WritingPrompts

Execute the following command (or run bash ./finetune.sh directly) to fine-tune HINT on ROCStories/WritingPromts:

cd ./model
env CUDA_VISIBLE_DEVICES=0 CUDA_LAUNCH_BLOCKING=1 python3.7 -u finetune.py \
    --data_dir ../Data/data/roc \ # ../data/data/wp
    --output_dir=./bart_post_bc_ft_roc \
    --save_top_k 80 \
    --train_batch_size=10 \
    --eval_batch_size=10 \
    --num_train_epochs 10 \
    --model_name_or_path ./bart_post_bc \
    --learning_rate=3e-5 \
    --fp16 \
    --gpus 1 \
    --do_train \
    --n_val 1000 \
    --val_check_interval 0.1 \
    --overwrite_output_dir \

You can add --sbert and --reorder to use the proposed two pretraining tasks as the auxiliary tasks for fine-tuning.

4. Generation and Computing Perplexity

Execute the following command to generate texts:

cd ./eval
device=cuda:0
model_name_path=../model/bart_post_bc_ft_roc
data_dir=../Data/ini_data/roc
env CUDA_VISIBLE_DEVICES=1 python3.7 ./gen.py $device $model_name_path $data_dir
env CUDA_VISIBLE_DEVICES=1 python3.7 ./ppl.py $device $model_name_path $data_dir

The generation results will be saved under the results directory.

5. Evaluation

Execute the following command to generate texts:

cd ./eval
python3.7 ./eval.py

You can change result_list in the script to decide the results you want to evaluate.

Citation

Please kindly cite our paper if this paper and the code are helpful.

@misc{guan2021long,
      title={Long Text Generation by Modeling Sentence-Level and Discourse-Level Coherence}, 
      author={Jian Guan and Xiaoxi Mao and Changjie Fan and Zitao Liu and Wenbiao Ding and Minlie Huang},
      year={2021},
      eprint={2105.08963},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}