Accurate architecture recovery is essential for the effective maintenance and evolution of modern software systems. While recent approaches have improved recovery precision through multidimensional fusion by combining lexical information with dependencies and directory structures, two major challenges remain. The first is the semantic gap introduced by lexical information, which seeks to infer design intent and code entity responsibilities from naming and comments but is too shallow to accurately capture the underlying functional intent. The second is the incomplete utilization of available information in fusion frameworks that compress heterogeneous information into single edge weights. We address these challenges with REFINE, a novel recovery method that integrates functional semantics, dependencies, and directory structures within a unified fusion framework. REFINE employs code models to capture functional intent, encodes each information source as a vector representation, and performs weighted embedding fusion before clustering entities into architectural components. Our evaluation, covering nine software systems with ground-truth architectures, investigates four key dimensions: baseline comparison, functional semantic contribution, fusion effectiveness, and clustering impact. Results show that REFINE achieves an average precision improvement of 27% over representative baselines. These findings highlight REFINE as a reliable and efficient solution for architecture recovery.