Integrating Conversational Agents and Knowledge Graphs Within the Scholarly Domain
The Problem:
Large language models (LLMs) have revolutionised question answering, including within the scientific domain. However, scientific question-answering remains significantly challenging for the current generation of LLMs due to their reliance on highly specialised concepts.
A promising solution:
Integrate LLMs with knowledge graphs of research concepts (also known as Scientific Knowledge Graphs), ensuring that responses are grounded in structured and verifiable information.
One approach consists of retrieving a relevant portion of the knowledge graph and providing it as input for the LLM. In this way, the LLM will use the structure knowledge of the graph to provide more accurate responses.
An alternative, and more effective, strategy uses LLMs to translate questions posed in natural language into SPARQL queries, enabling the retrieval of relevant data.
Reporting the state-of-the-art:
- Effective LLM-KG integration strategies
- Approaches to scientific question-answering
- Insights from Knowledge Media Institute’s research
- Achievable robustness in scientific (and other highly technical) question-answering.
- Current challenges and future directions for the field.