Semantic SLAM is essential for high-level scene understanding in robotics and augmented reality. Recent NeRF-based implicit SLAM methods provide efficient scene representations, but often suffer from insufficient semantic integration, high memory consumption, and suboptimal feature alignment. In this paper, we present COS-SLAM, a novel framework that enhances both spatial and semantic expressiveness through a Tri-Plane representation with coordinate attention. Furthermore, we introduce a lightweight Pixel-To-Line Transformer module for direct extraction of line features from RGB pixels, and propose a depth-guided truncated Gaussian sampling strategy to improve sampling efficiency. Extensive experiments on the Replica and ScanNet benchmarks demonstrate that COS-SLAM achieves geometry and appearance reconstruction comparable to or surpassing state-of-the-art SLAM methods. Notably, for semantic reconstruction, COS-SLAM achieves an mIoU exceeding 93% on the Replica dataset, more than 10% higher than the baseline.