Mueller matrix spectroscopy ellipsometry (MMSE) combined with machine learning method is a promising metrology tool for assisting high-volume production of complex transistor structures such as Gate-all-around (GAA). Conventional wide-spectrum MMSE cannot avoid measurement errors caused by large dispersions. In the meantime, the large amount of data in wide-spectrum reduces the metrology efficiency. In this work, a new metrology method is developed using the effects of near-field electric fields. First, two important transistor structures within the GAA architecture are used as metrology objects, including isotropic and anisotropic structures. The near-field electric fields and Mueller matrix spectra based on different dimensions are calculated using rigorous coupled wave analysis. Next, machine learning method has been used to compare the training time and test accuracy of the wide-spectrum (0.2 ~ 1.5 μm) with that of the sensitive spectrum (the bands where surface plasmon resonances or localized plasmonic resonances occur). The training time for the sensitive spectrum is shortened by about 60%, with the same test accuracy. The proposed method improves the metrology efficiency of complex transistor structures and reduces the dispersion error caused by wide spectrum, thus improving the reliability of high-density integration.