KRGP: knowledge-based response generation with persona

To create a personalized response, a generative model must take into account personal information about the user, question asked, and domain knowledge. Therefore, it is necessary to learn how to extract relevant information that will help the generative model to compose a response to the user. In this work, we propose to split the process into three stages: selection of relevant sentences from the textual knowledge base, selection of the most suitable sentences of the textual persona description taking into account the extracted knowledge, and response generation based on the knowledge and persona. We use the sentence Transformer and adapt the algorithm from the CLIP paper to obtain contextual sentence embeddings to extract the most relevant text spans from the knowledge base. We have found that the focal loss shows better results in tasks of binary classification of a persona using the FoCus imbalanced dataset as an example. We have also shown that text2text Transformer BART performs well in the tasks of conditional response generation in a dialog. This system achieved a state-of-the-art result at the leaderboard of The 1st Workshop on Customized Chat Grounding Personahttps://sites.google.com/view/persona-knowledge-workshop/home. The code for this work is available at https://github.com/dmitrymailk/deeppavlov_focus.