ParallelCVAE: A Parallel CVAE Mechanism for Multi-Turn Dialog Response Generation Model

Dialog generation in natural language processing, with the goal of generating natural responses based on conversation history, has gained significant attention. This technology is used in various real-world situations such as chatbots and customer support. Open-domain dialog systems, in particular, have become popular due to their commercial potential. However, ensuring contextual relevance remains a challenge due to insufficient feature representation. Dialogue generation models produce short, less informative, and repetitive responses as a result of the difficulty of dealing with limited contextual representation. Current models fail to sufficiently emphasize user utterances as a separate source of information, resulting in inadequate user-centric feature representation. Consequently, there is a need for architectures that better capture context and integrate user utterances with the dialogue history. Efforts to enhance dialog systems involve extracting more detailed features and exploring diverse architectural approaches. In this paper, we propose a conditional variational autoencoder (CVAE) model that focuses on user utterances. Unlike traditional CVAE models, our method enhances the information of latent vectors obtained independently by parallel prior and recognition networks for user utterances. The results from our proposed model demonstrate that focusing on user utterances leads to increased length and improved BLEU score values for estimated utterances.