ISEE: Interactive Semantic Enrichment for Database Fields
Abstract
LLM-based agents are increasingly deployed for data-related tasks, including sense-making, exploration, and retrieval. However, their performance depends heavily on the clarity and completeness of data semantics. In practice, many field descriptions remain ambiguous or incomplete, as essential context—such as the meaning of a customized field—often originates from users' domain knowledge and is rarely documented publicly. This gap restricts agent performance in downstream tasks like entity-linking. To address this, we introduce ISEE, a novel and comprehensive Interactive Semantic Enrichment system. Given a data field description, ISEE scores its quality, gathers domain knowledge, and collaborates with users to enrich semantics. Through user studies, automated user simulation, quantitative evaluation, and case studies, we demonstrate that ISEE significantly reduces cognitive load, improves description quality, and enhances downstream task performance.
Introduction
As LLM-based agents become integral to data workflows, their effectiveness is contingent on the quality of underlying data semantics. Yet, real-world databases often contain field descriptions that are vague, incomplete, or documented only in private domain knowledge. Such gaps can undermine agent-driven processes, particularly in tasks requiring precise data linking and retrieval. This paper presents ISEE, a system designed to bridge this semantic gap through interactive, user-guided enrichment.
System Overview
ISEE operates in three stages: it first evaluates the clarity of a given field description using a scoring mechanism, then identifies missing context by tapping into user domain knowledge, and finally collaborates with the user to iteratively refine and expand the semantics. The system is built to be lightweight and adaptable, supporting various data schemas and user expertise levels.
Evaluation and Results
We conducted a comprehensive evaluation involving a user study, automated user simulations, quantitative metrics, and detailed case analyses. Results demonstrate that ISEE dramatically reduces the cognitive burden on users, elevates the quality of field descriptions, and improves downstream task accuracy, particularly for entity-linking applications.
Conclusion
ISEE offers a scalable solution to a persistent problem in data management: the gap between stored semantics and user intent. By enabling interactive enrichment, it empowers users to contribute domain knowledge effectively, leading to more robust and reliable LLM-based data agents. Future work will explore broader integration with data governance frameworks and automated enrichment pipelines.
via ArXiv AI