Year: 2025
Type: conference-paper
Abstract: Answering scholarly hybrid questions requires access to bibliographic facts stored in structured data, such as a Knowledge Graph (KG) and textual information. Existing Scholarly Hybrid Question Answering (SHQA) approaches rely on retrieving KG triples and documents from the Wikipedia text corpus and prompt an LLM (Large Language Model) for answers. However, the retrieval is heavily keyword-based, introducing noise into the context. Furthermore, despite detecting the entities in the question, the models do not attempt any question analysis. Therefore, we propose a new SHQA system that employs a bridge-generate approach. During the bridge phase, our system recursively identifies entity-encapsulating phrases within the question and resolves the entities leveraging the underlying KGs. It then formulates assertion statements based on the resolved entities and their corresponding phrases. In the generation phase, the system auto-generates context guided by the question and the assertions. Finally, it returns an answer prompting an LLM with the generated context, the assertions, and the question. Our approach outperforms previous approaches, addressing the identified gaps.
Authors: Tilahun Abedissa Taffa, Ricardo Usbeck
Institutions: Leuphana University of Lüneburg, Universität Hamburg
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Funder: Deutsche Forschungsgemeinschaft
OpenAlex ID: W4411549646