Generative Fuzzy System for Sequence-to-Sequence Learning via Rule-Based Inference.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Fuzzy Systems

Background:

• Generative models (GMs), including large language models (LLMs), excel at creating new data but often lack interpretability and broader knowledge acquisition.
• Their complex, data-driven nature hinders understanding and control of outputs, limiting robustness and generalization.

Purpose of the Study:

• To enhance generative tasks by integrating data-driven and knowledge-driven mechanisms.
• To introduce a novel generative fuzzy system (GenFS) framework that combines GM deep learning with fuzzy system interpretability.
• To develop an end-to-end GenFS-based model for sequence generation.

Main Methods:

• Leveraging fuzzy systems, a classical modeling approach, to complement GMs.
• Proposing the GenFS framework integrating deep learning with fuzzy logic's term-based interpretability and dual-driven mechanisms.
• Developing the FuzzyS2S model for sequence generation tasks.

Main Results:

• Conducted test studies on 12 datasets across machine translation, code generation, and summary generation.
• FuzzyS2S demonstrated superior accuracy and fluency compared to the transformer model.
• Outperformed state-of-the-art models T5 and CodeT5 in specific application scenarios.

Conclusions:

• The GenFS framework effectively enhances generative models by incorporating knowledge-driven insights.
• FuzzyS2S offers a promising approach for more robust and interpretable sequence generation.
• This hybrid approach shows potential for advancing AI capabilities in complex generative tasks.