Interactive robot dialogue text fusion algorithm based on Transformer-XL model and knowledge distillation

To enhance the understanding of contextual relationships between sentences in interactive robot dialogue texts, reduce the computational cost of deploying dialogue systems, and improve their operational efficiency in resource-constrained environments, this study investigates a text fusion algorithm for interactive robot dialogue based on the Transformer-XL model and knowledge distillation. The algorithm employs a convolutional neural network (CNN) to extract local features from the dialogue text. By integrating two innovative mechanisms of Transformer-XL—the recurrence mechanism and relative positional encoding—into the Transformer architecture, the algorithm effectively captures global features that represent the contextual relationships among sentences in the dialogue text. After fusing the local and global features, the combined features are decoded into dialogue responses. Knowledge distillation is then applied to treat the above-described process as a complex teacher model, transferring its knowledge to a lightweight student model to improve both performance and computational efficiency. Experimental results show that the proposed algorithm can generate coherent, meaningful, and contextually appropriate responses according to user inputs and conversational context, thereby enhancing the interactive effectiveness of dialogue robots. The improvements to the Transformer model, along with the incorporation of knowledge distillation, help preserve contextual information when processing long sequences, leading to higher-quality conversational responses.