Dysarthria is commonly associated with impairments in lip and tongue muscle movements, which substantially reduce speech intelligibility and complicate both early diagnosis and rehabilitation assessment. Conventional clinical evaluation methods rely heavily on subjective rating scales or specialized measurement equipment, making it difficult to achieve both accuracy and practicality in real-world settings. To address this issue, we propose a video-based, fully sensor-free automated framework for assessing facial motor function, providing a more feasible solution for both clinical and remote rehabilitation scenarios. Specifically, we develop an attention-enhanced three-dimensional residual network (AE-3DResNet), in which a three-dimensional multi-scale attention module (3D-Efficient Multi-Scale Attention,3D-EMA) is integrated into the 3D ResNet-18 backbone to selectively enhance fine-grained spatiotemporal motion cues that are highly relevant to dysarthria assessment. In addition, we construct the first patient video dataset covering multi-level lip and tongue movement tasks, and alleviate its inherent long-tail distribution using a combination of Focal Loss and data resampling strategies. Experimental results show that the proposed framework achieves accuracies of 77.14% and 77.19% for three-level severity classification of lip and tongue motor dysfunction, corresponding to relative performance improvements of 6.66% and 7.89% over baseline models, respectively. This study not only provides a feasible technical pathway for objective and quantitative dysarthria assessment, but also features a deployment-friendly design that requires no additional hardware, showing strong potential for future clinical application and tele-rehabilitation.