Improving the Consistency of Dialog Models Through Speaker Separation Learning

In recent years, dialog systems, a type of application in the field of natural language processing, have become more prevalent in our daily lives, such as through help desk services. In dialog response generation, responses generated for a specific context may differ from those for other contexts not only grammatically but also semantically in some cases. Thus, simply applying translation technologies would cause issues with the diversity of the generated responses. Previous studies, such as VHRED and GVT, used sampled latent variables for response generation to achieve response diversity. In this study, we propose a method (extended GVTSC) for classifying dialogs before reflecting them in internal dialog processing, in addition to the characteristics of each speaker, to improve diversity while maintaining consistency.