Personalizing Dialogue Agents for Russian: Retrieve and Refine

Abstract

Currently, the development of automatic dialogue systems is in demand not only for traditional applications (increasing the level of automation of contact centers) but also for relatively new cases (development of virtual assistants, smart speakers, interactive robots). Adding information about characteristics of a person, which the agent must take into account when generating a response, i.e. personification of conversational agents, helps to increase user loyalty and engagement. This paper presents a study of Retrieve and Refine models for automatic generation of utterances of a personalized Russian-speaking dialogue agent. To train models in Russian, the Toloka Persona Chat Rus dataset is used. Refine models that used an adaptation of the BlenderBot model for the Russian language showed worse performance than for datasets in English. For Retrieve models, a solution based on the BERT encoder model was proposed, which made it possible to obtain the value of the metrics hits@1=0.705 for the model without a person, and hits@1=0.717 for the model with a person.