Cactus: Towards Psychological Counseling Conversations using Cognitive Behavioral Theory

This is the official GitHub repository for Cactus: Towards Psychological Counseling Conversations using Cognitive Behavioral Theory accepted at EMNLP Findings 2024.

Citation

@misc{lee2024cactus,
      title={Cactus: Towards Psychological Counseling Conversations using Cognitive Behavioral Theory}, 
      author={Suyeon Lee and Sunghwan Kim and Minju Kim and Dongjin Kang and Dongil Yang and Harim Kim and Minseok Kang and Dayi Jung and Min Hee Kim and Seungbeen Lee and Kyoung-Mee Chung and Youngjae Yu and Dongha Lee and Jinyoung Yeo},
      year={2024},
      eprint={2407.03103},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2407.03103}, 
}

Link

Our dataset & model are available here.

CACTUS Inference README

Setup

1. Virtual Environment Setup

We recommend you create a virtual environment using conda or virtualenv.

Using Conda

conda create -n therapy-session python=3.8
conda activate therapy-session

Using Virtualenv

# if virtualenv is not installed
pip install virtualenv

# Create a virtual environment
virtualenv .venv
source .venv/bin/activate  # Linux & macOS
.venv\Scripts\activate  # Windows

2. Installing Required Packages

After activating the virtual environment, install the necessary packages using the requirements.txt file.

pip install -r requirements.txt

3. Configuring the Settings File

Copy the config.yaml.example file in the conf.d folder to create a config.yaml file. Then, fill in the following content in the config.yaml file.

openai:
  key: <<Your openai API key>>

llama2:
  host: http://<<Server IP or URL>>/v1

llama3:
  host: http://<<Server IP or URL>>/v1

Adding a Counselor Agent

To add a counselor agent, follow these steps.

1. Creating a Prompt File

• The prompt file should be located in the prompts folder.

• The file name pattern should follow the format agent_{counselor_type}_{llm_type}.txt. Example: agent_cactus_chatgpt.txt

• The prompt file should include a template for generating the counselor's response.

  Client information: {client_information}
  Reason for counseling: {reason_counseling}
  CBT plan: {cbt_plan}
  History: {history}
  

2. Adding a New Counselor Agent Class

Create a new counselor agent class by inheriting from the CounselorAgent class. Ensure to set self.language to either english for English or chinese for Chinese.

class NewCounselorAgent(CounselorAgent):
    def __init__(self, llm_type):
        super().__init__(llm_type)
        self.language = "english"  # For English
        # self.language = "chinese"  # For Chinese
        prompt_text = self.load_prompt(f"agent_new_{llm_type}.txt")
        self.prompt_template = PromptTemplate(
            input_variables=["history"],
            template=prompt_text)

    def generate(self, history):
        # Override the generate function if necessary
        history = '\n'.join(
            [
                f"{message['role'].capitalize()}: {message['message']}"
                for message in history
            ]
        )
        prompt = self.prompt_template.format(history=history)
        return self.llm.generate(prompt)

3. Adding the New Counselor to LLMFactory

Add the new counselor agent to the LLMFactory class.

class LLMFactory:
    @staticmethod
    def get_llm(llm_type):
        if llm_type == "chatgpt":
            return ChatGPT()
        elif llm_type == "llama2":
            return LLama2()
        elif llm_type == "llama3":
            return LLama3()
        elif llm_type == "new":
            return NewCounselorAgent(llm_type)
        raise ValueError(f"Unsupported LLM type: {llm_type}")

Adding a New LLM

To add a new LLM, follow these steps.

1. Creating a New LLM Class

Create a new LLM class by inheriting from the LLM abstract class.

class NewLLM(LLM):
    def __init__(self):
        config = get_config()
        api_key = config['new']['key']
        self.llm = OpenAI(
            temperature=0.7,
            model_name="new-model",
            openai_api_key=api_key
        )

    def generate(self, prompt: str) -> str:
        response = self.llm.invoke(prompt)
        return response.content

2. Adding the New LLM to LLMFactory

Add the new LLM to the LLMFactory class.

class LLMFactory:
    @staticmethod
    def get_llm(llm_type):
        if llm_type == "chatgpt":
            return ChatGPT()
        elif llm_type == "llama2":
            return LLama2()
        elif llm_type == "llama3":
            return LLama3()
        elif llm_type == "new":
            return NewLLM()
        raise ValueError(f"Unsupported LLM type: {llm_type}")

Running Counseling-Eval

1. Prepare necessary files and folders

• Ensure the necessary prompt files are available in the prompts folder.
• The input file should be a JSON file containing the client intake form.

2. Run the program

Run the program using the following command.

python script.py --input_file {path to input file} --output_dir {output directory} --counselor_type {counselor type} --llm_type {LLM type} --max_turns {maximum number of turns}

Example:

python script.py --input_file ./data/intake_forms.json --output_dir ./output --counselor_type cactus --llm_type chatgpt --max_turns 20

3. Using the Execution Script

You can use the 'scripts/inference.sh' script for easy execution. Run it as follows:

sh scripts/inference.sh

4. Running the vLLM Server

All models except chatgpt (such as llama2, llama3, etc.) need to run on the vLLM server. Refer to the scripts/run_vllm.sh script for this.

sh scripts/run_vllm.sh

This script includes all the commands necessary to set up and run the vLLM server. With the vLLM server running, you can simulate the counseling session using the program.