Oral motor exercises are important for people of all ages, especially the elderly, as they help maintain optimal oral health, including activities such as eating, swallowing, and speaking. Visually recognizing oral movements can be challenging due to subtle variations in individual facial expressions. Furthermore, even when accurately recognized, judging the effectiveness of these exercises can be difficult. To address these challenges, this paper proposes to utilize machine learning models for recognition and then evaluate oral motor exercises using a Mamdani fuzzy inference system. The Wang-Mendel method has been shown to be effective in generating interpretable fuzzy systems by generating fuzzy rules directly from numerical data. However, it must be acknowledged that the main application of rules generated by this method is for recognition purposes rather than for exercise evaluation. In response to this limitation, this paper proposes a new method to filter out effective variables from Wang-Mendel method rules and use these variables to generate inference rules for the Mamdani fuzzy inference system for the assessment task. In order to evaluate the effectiveness of the proposed method, 13 different oral motor exercise datasets of 9 college students were collected. Experimental results demonstrate the effectiveness of the proposed method in rejecting unlearned exercises and recalibrate predictions, achieving an accuracy of 73% under leave-one-subject-out cross validation.