| Summary: | Personalized dialogue agents are capable of generating responses consistent with a specific persona. Typically, personalized dialogue agents generate responses based on both the dialogue history and a representation of the agent’s desired persona. As it is impractical to obtain the persona representations for every interlocutor in real-world implementations, recent works have explored the possibility of generating personalized dialogue by finetuning the agent with dialogue examples corresponding to a given persona instead. However, in real-world implementations, a sufficient number of corresponding dialogue examples are also rarely available. Hence, in this paper, we introduce the Dual Latent Variable Generator (DLVGen), a variational personalized dialogue agent capable of generating personalized dialogue without any persona information or any corresponding dialogue examples. Unlike previous works, DLVGen models the latent distribution over potential dialogue response intents as well as the latent distribution over the agent’s potential persona. During inference, latent variables are sampled from both distributions and fed to the decoder. Extensive experiments on the popular ConvAI2 personalized dialogue corpus show that DLVGen is capable of generating natural, persona consistent responses. Additionally, we also introduce a variance regularization and response selection approach which further improved overall response quality.
|
|---|