Counterspeeches up my sleeve! Intent Distribution Learning and Persistent Fusion for Intent-Conditioned Counterspeech Generation

The code and dataset for our ACL'23 paper: Counterspeeches up my sleeve! Intent Distribution Learning and Persistent Fusion for Intent-Conditioned Counterspeech Generation

The paper proposes the novel task of intent-specific counterspeech generation. We develop IntentCONAN, a diverse intent-specific counterspeech dataset by augmenting the existing CONAN dataset. We propose QUARC, a two-stage framework for generating intent-conditioned counterspeeches by fusing the learned vector-quantized representations of each intent category.

Repository Structure

The structure of this repository is as follows:

Folders

• configs/: contains the YAML configuration files for the model parameters (for fast prototyping).
• data/: Contains the training, validation and testing datasets.
• modeling/: contains the model architectures and utilities.
• models/: folder to store the trained models
• utils/: contains various utilities used for training and testing

Files

• process_data_clime.py: contains data processing for clime
• process_data_cogent.py: contains data processing for cogent
• train_clime.py: script for training clime
• train_cogent.py: script for training cogent
• results.py: contains the evaluation script to obtain the results
• requirements.txt: the required libraries with their versions

Running the code

1. To run the code, first install the necessary requirements via the following command:

   pip install requirements.txt
   

2. Then, set the hyperparamters by creating a configuration file in the config folder. A sample configuration file is provided, copy it and change the desired parameters.

3. Run CLIME by executing the following command:

   python train_clime.py --config_path configs/<your_config>.yaml
   

4. Run COGENT by executing the following command:

   python train_cogent.py --config_path configs/<your_config>.yaml
   

Contact

In case of any queries, please reach out to Rishabh Gupta.