SummaryGPT: Leveraging ChatGPT for Summarizing Knowledge Graphs

Abstract

Semantic summaries try to extract compact information from the original knowledge graph (KG) while reducing its size for various purposes such as query answering, indexing, or visualization. Although so far several techniques have been exploited for summarizing individual KGs, to the best of our knowledge, there is no approach summarizing the interests of the users in exploring those KGs, capturing also how these evolve. SummaryGPT fills this gap by enabling the exploration of users’ interests as captured from their queries over time. For generating these summaries we first extract the nodes appearing in query logs, captured from a specific time period, and then we classify them into different categories in order to generate quotient summaries on top. For the classification, we explore both the KG type hierarchy (if existing) and also a large language model, i.e. ChatGPT. Exploring different time periods enables us to identify shifts in user interests and capture their evolution through time. In this demonstration we use WikiData KG in order to enable active exploration of the corresponding user interests, allowing end-users to visualize how these evolve over time.