Combining Knowledge Graphs and Language Models to Answer Questions over Tables

Tables remain a primary modality for organizing and presenting information to people. We interact every day with Excel sheets, CSV files, tables in PDF documents, and web tables. Providing a natural language interface to query table information is paramount for several use cases. This demo shows a solution to query semantically described tables using natural-language questions. Our solution employs knowledge graphs as a medium to integrate tables coming from heterogeneous sources. Then, a transformer-based language model analyzes a user’s question and finds the answer in the semantically represented tables. During the demo session, we will show a use case developed in collaboration with DATEV eG, where tax consultants can efficiently query information from financial tables. Attendees will experience how a natural-language interface speeds up the information retrieval process from tables. They will also be allowed to ask their questions to a prepared dataset, showing the scalability of our solution. The video demo is available at https://owncloud.fraunhofer.de/index.php/s/uXFmUfzCta70rqN.