DisCup

This repository contains code for the paper DisCup: Discriminator Cooperative Unlikelihood Prompt Tuning for Controllable Text Generation which is appeared at EMNLP2022. If you have any questions, please feel free to create an issue or contact the email of the first author: zhanghanqing@bit.edu.cn

Welcome to our new work, Controllable Text Generation with Residual Memory Transformer (RMT), which can be regarded as a updated version of DisCup.

Description of Main files

• discrimination.py: discriminator training(i.e., detoxic classifier and sentiment classifier), and pure disriminator-based(FUDGE) generation
• prompt_tuning.py: the implemetation of vanilla prompt-tauning; it contains the vanilla prompt training and prompt-based generation
• distill_tuning.py: the implemetation of DisCup; it contains the discriminator cooperative unlikelihood prompt training and prompt-based generation
• /script: it contains the bash commands for model trainng and controllable text generation
• evaluate.py: evaluate the generated texts (i.e., dist1/dist-2/dist-3, Perplexity, and domain keyword coverage) with .txt format.

Dependence

• Install the following Conda environment
• • our code is bulit on python3.6
• • pip install -r requirements.txt
• Download the datasets: click here
• Downlad the check_points: click here
• Prepare the GPT2 models:
• • GPT2-small
• • GPT2-large

After downloading the trained check-points, you also can directly jump to Controllable Text Generation, conducting text generation experiments.

Discriminator Training

It contains the training process of attribute-discriminator, and it is the premise of the DisCup.

Sentiment classifer training

• cd ./script
• bash train_sentiment_disc.bash

Detoxic classifer training

• cd ./script
• bash train_detoxic_disc.bash

Parameter Configuration

• --data_path: training corpus data for classifer training
• --model_name_or_path： the path for the pretrained language model, we use GPT2-samll here
• --out_dir: the output directory to save the check-point
• --template: configure the prompt length of the control-prompt

Control-prompt Tuning

It contains the training process of control-prompts for vanilla-prompt tuning and DisCup.

Sentiment control task

• cd ./script

• bash train_sentiment_distill.bash or bash train_sentiment_prompt.bash

Detoxic task

• cd ./script

• bash train_detoxic_distill.bash or bash train_detoxic_prompt.bash

Parameter Configuration

• --data_path: the training corpus for prompt-tuning, attribute-specific corpus should be set for vanilla prompt-tuning
• --model_name_or_path： the path for the pretrained langauge model, we use GPT2-Large here
• --out_dir: the output directory to save the control-prompts
• --disc_embedding_checkpoint: the path of trained discriminators, it only needs to be specified in DisCup
• --template: configure the prompt length
• --ranking_scope: configure the size of re-ranked candidate tokens, it only needs to be specified in DisCup
• --corpus_type: the attribute of control-prompts, it contain [positive" "negative"] in sentiment control generation, only 'positive' is optional in toxicity avoidance task
• --temperature: configure the distribution shapeness of re-ranked candidate tokens, it only needs to be specified in DisCup

Controllable Text Generation

It contains the generation processes for vanilla-prompt and DisCup.

Sentiment control task

• cd ./script

• bash generate_sentiment_distill.bash or bash generate_sentiment_prompt.bash

Detoxic task

• cd ./script

• bash generate_detoxic_distill.bash or bash generate_detoxic_prompt.bash

Parameter Configuration

• --data_path: the prompts data for text generation
• --model_name_or_path： the path for the pretrained langauge model, we use GPT2-Large here
• --file_name: the output directory to save the generation results, it is saved with '.csv' format
• --embedding_checkpoint: the path of the saved control-prompts
• --template: configure the prompt length, which is consistent to the actual prompt length of embedding_checkpoint
• --prompt_type: specify the prompt type, it contain ["neutral" "positive" "negative"] in sentiment generation, only 'negative' is optional in detoxic
• --target_type: consistent to the type embedding_checkpoint it contain [positive" "negative"] in sentiment control generation, only 'positive' is optional in toxicity avoidance task

Evaluation:

• For sentiment control
  you can refer to sentiment_classifier.ipynb, and evaluate the correctness and PPL of your generated csv file. Then, you can convert the csv file to txt, using the function save_csv_to_text, so as to run the evaluate.py to obtain the results of dist-1/2/3 and and domain keyword coverage.

• For toxicity avoidance
  you could test the toxicity prob with the Google API, the measurement of dist-1/2/3 and PPL is same as sentiment control task.

Citation

@inproceedings{zhang-song-2022-discup,
    title = "{D}is{C}up: Discriminator Cooperative Unlikelihood Prompt-tuning for Controllable Text Generation",
    author = "Zhang, Hanqing  and
      Song, Dawei",
    booktitle = "Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing",
    month = dec,
    year = "2022",
    address = "Abu Dhabi, United Arab Emirates",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.emnlp-main.223",
    pages = "3392--3406",
    abstract = "Prompt learning with immensely large Casual Language Models (CLMs) has been shown promising for attribute-controllable text generation (CTG). However, vanilla prompt tuning tends to imitate training corpus characteristics beyond the control attributes, resulting in a poor generalization ability. Moreover, it is less able to capture the relationship between different attributes, further limiting the control performance. In this paper, we propose a new CTG approach, namely DisCup, which incorporates the attribute knowledge of discriminator to optimize the control-prompts, steering a frozen CLM to produce attribute-specific texts. Specifically, the frozen CLM model, capable of producing multitudinous texts, is first used to generate the next-token candidates based on the context, so as to ensure the diversity of tokens to be predicted. Then, we leverage an attribute-discriminator to select desired/undesired tokens from those candidates, providing the inter-attribute knowledge. Finally, we bridge the above two traits by an unlikelihood objective for prompt-tuning. Extensive experimental results show that DisCup can achieve a new state-of-the-art control performance while maintaining an efficient and high-quality text generation, only relying on around 10 virtual tokens.",
}

The part of the code was built on top of All NLP Tasks Are Generation Tasks: A General Pretraining Framework.