Background <p>Traditional language classrooms, due to the absence of an emotional feedback mechanism, are prone to an imbalance between learners' cognitive and emotional engagement, affecting the efficiency of knowledge internalization. With the development of artificial intelligence technology, intelligent educational robots have gradually evolved from knowledge transfer tools into emotional interactive agents. However, the systematic driving effect of their emotional interaction mechanism on language acquisition remains unclear, especially in the context of basic education, where the dynamic patterns and group adaptability of technological effects urgently need to be explored. Current research focuses on the cognitive assistance and emotional interaction functions of educational robots. However, there are still controversies regarding the fidelity and ethical boundaries of affective computing technology, and multimodal data fusion faces technical bottlenecks such as cross-modal temporal alignment and dynamic weight optimization. In addition, differences in educational resources between urban and rural areas may restrict the universality of emotional interaction technology. Existing research is mostly limited to single-scenario validation and lacks large-sample stratified research support. </p> Methods <p>This study adopts a mixed longitudinal experimental design, deploys a multimodal data collection system in a natural classroom environment, and constructs a dynamic weight allocation model to optimize the accuracy of emotion recognition. By stratified sampling from urban and rural areas, 240 primary and secondary school students were selected to compare the language acquisition effects of the experimental group and the control group. Combining the "emotion-cognition-behavior" ternary integration model, the moderating effects of technological acceptance and classroom ecology were analyzed.</p> Findings <p>The emotional interaction mechanism of intelligent educational robots significantly enhances language acquisition efficiency through multimodal fusion technology, and the dynamic weight model enables emotional recognition accuracy to surpass the limitations of traditional algorithms.</p> Conclusions <p>Younger learners are more sensitive to anthropomorphic interactions, while logical reasoning support needs to be strengthened during junior high school. The urban-rural differences reveal that technological deployment needs to be coordinated with teacher training and infrastructure optimization.</p>