Causal Inference for Human-Language Model Collaboration (NAACL 2024)

This is the code repository for our NAACL 2024 paper, "Causal Inference for Human-Language Model Collaboration."

Run the Model

You have the option to run the model with or without style extractions.

With Style Extractions

To enable style extractions, you can set the following parameters:

python main.py --decompose_a=1 --decompose_a_model=CVAE  
# For using CVAE (or PCA) for style extraction

Without Style Extractions

Simply run the model without setting the --decompose_a and --decompose_a_model flags.

python main.py

The results include performances on both observational and counterfactual data with or without G-estimation.

Data Options

You can specify the dataset using the --data_name flag. The available datasets are:

• coauthor: Designing a Human-AI Collaborative Writing Dataset for Exploring Language Model Capabilities.
• baize: An Open-Source Chat Model with Parameter-Efficient Tuning on Self-Chat Data.
• dialcon: Human machine collaboration approaches to build a dialogue dataset for hate speech countering.

---