LLMs Show Limits in Conceptual Database Modeling
A recent study published on arXiv investigates the capability of Large Language Models (LLMs) in creating Entity-Relationship diagrams from natural language inputs. The research involved three distinct LLMs and employed techniques including Zero-Shot, Chain of Thought, and Chain of Thought with a Verifier, assessing their performance on varying task difficulties. Findings indicated that while the models performed adequately on simpler tasks, they encountered notable difficulties with more intricate conceptual modeling scenarios, highlighting the current limitations of LLMs in this area of knowledge representation and diagram generation.
Key facts
- Study analyzes LLMs for automatic ER diagram generation
- Three LLMs tested with three prompting techniques
- Scenarios had progressively increasing complexity
- Qualitative analysis compared diagrams to textual requirements
- LLMs performed reasonably in less complex scenarios
- Published on arXiv with ID 2605.11986
- Focus on structural and semantic adherence
- Prompt engineering techniques included Zero-Shot, Chain of Thought, and Chain of Thought + Verifier
Entities
Institutions
- arXiv