A dual-dimension collaborative enhancement framework to boost language model spatial semantic understanding

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Cognitive Science

Background:

• Spatial expressions are crucial for understanding language, requiring linguistic, cognitive, and world knowledge.
• Spatial semantic understanding is challenging for small language models due to limited reasoning capabilities.
• Existing methods struggle with the complex logical reasoning needed for spatial semantics.

Purpose of the Study:

• To develop a framework that enhances the spatial semantic understanding of small language models.
• To enable small language models to approximate the performance of large language models in spatial reasoning.
• To improve performance in low-resource scenarios for natural language understanding tasks.

Main Methods:

• A cognition-data collaborative enhancement framework is proposed.
• Chain-of-thought is injected to decompose reasoning into transferable cognitive units.
• Semi-supervised learning with sequence confidence extracts high-quality spatial relationship data from unlabeled text.

Main Results:

• The framework enables small language models to achieve performance comparable to large language models in spatial semantic reasoning.
• Significant performance improvements were observed for small language models in low-resource settings.
• The synergistic approach of cognitive guidance and data integrity forms an effective closed loop.

Conclusions:

• The proposed framework offers a novel paradigm for semantic understanding in resource-constrained environments.
• It effectively bridges the performance gap between small and large language models in complex reasoning tasks.
• This approach facilitates more capable natural language understanding with smaller, more efficient models.