Neuro-symbolic NLP: taxonomy, assessment, and directions

Area of Science:

• Natural Language Processing (NLP)
• Artificial Intelligence (AI)
• Computational Linguistics

Background:

• Purely neural and symbolic approaches to NLP have inherent limitations.
• Neuro-symbolic (NeSy) systems offer a promising hybrid solution.
• Existing NeSy systems lack a unified classification framework.

Purpose of the Study:

• To survey recent developments in NeSy NLP.
• To propose a novel classification framework for NeSy systems.
• To analyze the performance and potential of different NeSy architectures.

Main Methods:

• Developed a taxonomy combining Kautz's integration types and Lappin's injective-federative distinction.
• Assessed system performance on benchmarks for compositional generalization, reasoning, and robustness.
• Examined existing efforts in integrating linguistic theory into NeSy models.

Main Results:

• Federative NeSy architectures consistently outperform injective approaches.
• Federative architectures remain significantly underexplored in current research.
• The proposed taxonomy can guide the development of future NeSy architectures.

Conclusions:

• Federative NeSy systems hold significant potential for advancing NLP.
• Further research is needed to scale federative systems and achieve tighter integration.
• Future architectures should leverage formal linguistic frameworks for enhanced capabilities.