Neuro-symbolic NLP: taxonomy, assessment, and directions

Neuro-symbolic (NeSy) approaches promise to overcome the limitations of purely neural and purely symbolic NLP. In this paper we survey recent development in NeSy NLP and propose a system classification framework that combines Kautz's integration types with Lappin's injective-federative distinction. We then apply this taxonomy and show that even though federative architectures consistently outperform injective approaches, they remain underexplored. We assess performance across compositional generalization, reasoning, and robustness benchmarks, examine existing attempts at linguistic theory integration, and show how the taxonomy can guide future architectures that preserve the strengths of formal linguistic frameworks. We conclude with directions for scaling federative systems and exploring tighter integration.