Progressive Generation

Code to be further cleaned up soon...

This repo contains preliminary code of the following paper:

Progressive Generation of Long Text
Bowen Tan, Zichao Yang, Maruan AI-Shedivat, Eric P. Xing, Zhiting Hu
pre-print, 2020

The current code includes implementation of:

• the progressive generation model
• extensive evaluation metrics:
  • Frechet Bert Distance (Montahaei et al.)
  • MS-Jaccard (Montahaei et al.)
  • Forward-Backward BLEU (Shi et al.)
  • TF-IDF Feature Distance (proposed in our paper)

Requirements

torch==1.2.0
transformers==2.5.1
fairseq==0.9.0

It needs at least 4 GPUs on your device if you want to finetune GPT2-Large baseline, otherwise 2 GPUs are enough.

Download Data

python download/download_cnn.py
python download/download_writing_prompts.py

Train

python train.py \
    --dataset [cnn/wp] \
    --prog_steps null-{...}-full \
    --first_model [bart/gpt2/gpt2-large]

• --first_model specifies the type of the first-stage model.

The training log will be stored in training_logs/{your setting}/:

• training_logs/{setting}/log.txt: evaluation loss of each checkpoint.
• training_logs/{setting}/ckpt_gens/step{}.txt: ~10 generation examples on dev set of each checkpoint.
• training_logs/{setting}/best_model.pt: best checkpoint model according to evaluation loss.

Check scripts/train_all.sh for all commands for training.

Generate

python generate.py \
    --dataset [cnn/wp] \
    --prog_steps null-{...}-full \
    --first_model [bart/gpt2/gpt2-large]

Generated texts will be stored in generated_texts/{your setting}/:

• generated_texts/{setting}/gen.txt: generation log.
• generated_texts/{setting}/gen.pickle: all generated texts stored into a pickle file.

Check scripts/gen_all.sh for all commands for generation.

Evaluate

python evaluate.py \
    --dataset [cnn/wp] \
    --prog_steps null-{...}-full \
    --first_model [bart/gpt2/gpt2-large]

Check scripts/eval_all.sh for all commands for evaluation.

Present Results

python present_eval_results.py \
    --dataset [cnn/wp] \
    --metric [ms_jaccard/frechet_bert_distance/tfidf_distance/forward_backward_bleu]

Check scripts/present_all.sh for all commands for presenting.

Results

For the simplicity to run our code, we reduce the test set to 1K examples (numbers in our paper are from 5K examples). If you want to change it back to 5K, see download/download_cnn.py and download/download_writing_prompts.py.

Results of 1K examples are as below. The superiority of our progressive models is already significant.

Dataset: CNN

Dataset: WritingPrompts