Abstract: To address the semantic ambiguity and entity boundary fuzziness arising from the high expressional flexibility, insufficient terminology standardization, and strong contextual dependencies in texts from the transmission line construction domain during the digital transformation of new power systems, this paper proposes an innovative recognition framework integrating bio-inspired computing and adaptive feature selection. The framework leverages a bio-inspired computing module based on a bidirectionally enhanced nonlinear spiking neural system, empowering the model with deep modeling capabilities for complex nonlinear semantic patterns to effectively resolve terminological ambiguity. Concurrently, it incorporates an adaptive feature selection mechanism using a gated residual temporal convolutional network, which precisely captures long-range contextual dependency features through dynamic gating and residual structures. These two modules extract complementary features via a deep synergy mechanism, significantly enhancing boundary accuracy and semantic robustness for entity recognition in generative texts. This provides scalable technical underpinning for application scenarios such as intelligent knowledge base construction for transmission line construction.