DBIS

Entity resolution is a prerequisite for any unified scholarly knowledge graph. The problem is well-studied in isolation, with established algorithmic solutions for string matching, clustering, and identifier lookup. Yet in practice, scholarly metadata presents a layered difficulty: trivial cases (exact ORCID match) coexist with genuinely ambiguous ones (two “W. Meier” at institutions whose names differ across platforms) within the same dataset. Applying expensive techniques uniformly, whether LLM-based reasoning or exhaustive API cross-referencing, wastes resources on cases that a simple join would resolve, while cheap heuristics alone fail silently on the hard cases.
The Dagstuhl Research Online Publication Server (DROPS) illustrates this directly. Its JSON-LD metadata defines authors locally within each document: the same person may appear with varying name forms, different affiliation strings, and inconsistent ORCID coverage across files. Some cases resolve trivially via a shared ORCID; others require querying external APIs (OpenAlex, CrossRef, ROR) and reasoning about conflicting evidence, for instance when OpenAlex merges two distinct people or when an affiliation string does not match any ROR entry without normalization. This research will design, implement, and evaluate a layered entity resolution architecture that routes each candidate pair through progressively more expensive resolution stages: fast deterministic matching first, algorithmic similarity scoring second, and LLM-based agentic reasoning only for the residual ambiguous cases. The central question is whether this layered strategy can achieve resolution quality comparable to a full agentic approach at a fraction of the computational cost.

Research Questions

• How can a layered entity resolution architecture be designed that routes candidate pairs through progressively expensive stages and what routing criteria determine when a case should be escalated to the next layer?
• What resolution strategies should the agent employ when external sources provide conflicting or overlapping identity sets, and how does the agent’s reasoning-based reconciliation compare to traditional deterministic or ML-based matching approaches?
• How effective is the agentic approach at resolving entities where key attributes (particularly affiliations) are unresolved or inconsistently represented across platforms, requiring multi-step verification chains?
• What is the cost–quality trade-off of the layered approach compared to (a) a purely algorithmic baseline and (b) a full agentic approach applied uniformly, measured in resolution quality (precision/recall), computational cost (API calls, LLM inference tokens), and end-to-end latency?