Blessing-Generation

Dataset and Models for GEM 2022 Workshop

"Towards Attribute-Entangled Controllable Text Generation: A Pilot Study of Blessing Generation".

Data

The full processed dataset is placed in resources/all_data.csv.

GPT2, T5 Models and Evaluation

Requirements

• torch == 1.10.0
• transformers == 4.14.1
• pytorch-lightning == 1.6.2
• scikit-learn == 1.0.2
• nltk == 3.6.3

Prepare Data

The overall dataset (csv file) should be divide into two files train.csv and test.csv. Put them in the resources directory.

Train

Run the following command to train a blessing generation model:

python -u main.py \
    --output_path output \
    --model_name gpt2 \
    --max_length 200

The model name gpt2 can be replaced with t5-base for training T5 model.

After each training epoch, the checkpoint will be saved in the output directory.

Test

Run the following command to test a model:

python -u main.py \
    --output_path output \
    --model_name gpt2 \
    --max_length 200 \
    --mode test \
    --ckpt_path <ckpt_file> # <ckpt_file> should be replaced with the specific filename of the checkpoint

The generated blessing sentences will also be output in the output directory in json format.

Evaluation

Modify some variables related to the directory and filename in metrics.py and execute it to calculate the metrics.

The metrics are saved in some csv files in the same directory as the prediction file, i.e., output.

GPT2 + CVAE and GPT2 + Adv.

The code is directly borrowed from TransformerCVAE.

Prepare Data

Execute write_data_to_cvae.py to convert the csv files train.csv and test.csv mentioned in the last section to the input data for CVAE and adversarial architecture.

Train and Test

Run the following commands for training the two models.

cd CVAE
python train_adv.py # For training GPT2 + Adv.
python train_cvae.py # For training GPT2 + CVAE

During the training process, a checkpoint and a text file containing prediction result in the test dataset will be output every few training rounds.

Evaluation

Modify and execute vae_postprocessing.py to convert the prediction results to a json file for calculating the metrics.

Follow the Evaluation part of the last section to obtain the all metrics.