Using graph and text knowledge bases in Dream dialogue assistant

The paper describes components of DREAM dialogue assistant for detection of entities in the user utterance, entity linking and extraction of facts from the knowledge base for a response generation. The entity extraction component finds substrings which correspond to entities in the user utterance. Entity substrings are linked to Wikidata ids and Wikipedia page titles, which are used in queries to knowledge bases for extraction of facts, viz. Wikidata triplets and Wikipedia page paragraphs. If the user utterance is a factoid question, knowledge base question answering and open-domain question answering components find answer in extracted facts. Wikipedia paragraphs are utilized in the scenario of DFF Wiki Skill for discussion of entities the user is interested in and in the skill based on the generative model (Knowledge Grounding Skill). Knowledge bases in the dialogue assistant allow us to analyze the user utterance and generate interesting and meaningful responses.