Reasoning over Uncertain Text by Generative Large Language Models

This repository is dedicated to the research and findings presented in the paper "Reasoning over Uncertain Text by Generative Large Language Models." Our study introduces the Bayesian Linguistic Inference Dataset (BLInD) and delves into the capabilities and limitations of Large Language Models (LLMs) in executing probabilistic reasoning with explicitly quantified uncertainties.

Dataset Overview

The BLInD dataset is crafted to evaluate the probabilistic reasoning skills of LLMs, featuring:

• A foundational Bayesian Network for each instance.
• A textual description detailing the structure of the Bayesian Network.
• Probabilistic queries posed in natural language.
• Precise answers corresponding to these queries.

This dataset is produced through a systematic pipeline (shown in the Figure above) that constructs Bayesian Networks, populates Conditional Probability Tables (CPTs), formulates queries, and translates all components into natural language. The dataset files and the Python scripts used for generation are available in the datasets directory.

Methodological Approaches

Examples of methods are shown in the Figure. Each method forms a prompt that starts with an instruction (purple boxes) that describes the problem and the solution format. Then, the context and query and answer that solves the query based on the context are demonstrated (based on our first in-context example) to the LLM. If NE or GG subtasks are used with the method, their instruction and answers are appended, as shown in the PAL method for NE and the MC method for GG.

Use the following links to access the Python code to execute and test these methods:

1. Baselines:
   • Basic Question Answering (BQA)
   • Chain of Thought (COT)
2. Subtasks:
   • Number Extraction (NE)
   • Graph Generation (GG)
3. Symbolic Methods:
   • Python code (PAL - Program Aided Language Models)
   • Monte Carlo algorithms (MC)
   • ProbLog

Repository Structure

• datasets/: Hosts the BLInD dataset and the Python scripts for its creation.
• BQAandCOT/: Contains scripts for testing BLInD with BQA and COT methodologies.
• PALandMC/: Includes scripts for employing PAL and MC techniques on BLInD.
• Problog/: Features scripts for applying ProbLog on BLInD.
• NE/: Provides scripts for the Number Extraction subtask.
• GG/: Offers scripts for the Graph Generation subtask.

Dependencies

Ensure you have Python 3.10.6 or newer installed. To install the required Python packages, run the following commands in your terminal:

python -m pip install --upgrade pip
pip install -r requirements.txt

Citation

To cite our work, please use the following BibTeX entry:

@misc{nafar2024probabilistic,
      title={Probabilistic Reasoning in Generative Large Language Models},
      author={Aliakbar Nafar and Kristen Brent Venable and Parisa Kordjamshidi},
      year={2024},
      eprint={2402.09614},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}