Large Language Model Enhanced Logic Tensor Network for Stance Detection

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Social media analysis requires accurate stance detection for understanding public opinion.
• Traditional deep learning models for stance detection face challenges with data requirements, interpretability, and domain knowledge integration.

Purpose of the Study:

• To introduce the First-Order Logic Aggregated Reasoning (FOLAR) framework for explainable and effective stance detection.
• To address limitations of traditional deep neural networks in stance detection tasks.

Main Methods:

• FOLAR integrates first-order logic (FOL) with large language models (LLMs).
• Key components include Knowledge Elicitation (using chain-of-thought prompting for FOL rules), Logic Tensor Network (LTN) for rule encoding, and Multi-Decision Fusion for robustness.
• The framework aims to enhance interpretability and incorporate human intentions.

Main Results:

• Experiments on standard benchmarks demonstrate FOLAR's effectiveness.
• The framework shows improved accuracy and interpretability in stance detection.
• FOLAR offers a robust solution by minimizing biases through aggregated decision-making.

Conclusions:

• FOLAR presents a novel and promising approach for explainable stance detection.
• The integration of FOL and LLMs enhances the efficacy and transparency of opinion mining.
• The public release of source code aims to facilitate further research in this domain.