The primary objective of this thesis is to develop a solution for one-to-many structural mapping suggestions, where fields from a source system are matched to potential candidate fields in a target system. A secondary use case is the detection of duplicate records, which builds on the same core similarity framework.

The goal is to improve upon a prototype by designing, implementing, and evaluating a similarity scoring method for record-level matching in structured datasets. The method will support configurable aspects, each combining selected attributes (e.g. field name, description, value samples), similarity signals (e.g. lexical, embedding-based), and weights. This configuration is intended to make the approach adaptable to different domains and matching scenarios.

The work includes implementing the core scoring method, re-implementing the backend logic and configuration, researching and incorporating similarity signals, defining evaluation datasets that represent various error classes, and optimizing hyperparameters for them. Controlled perturbations will be applied to assess resistance to noise. The final deliverable is a validated core matching method suitable for integration into broader data processing solutions, developed in cooperation with Natuvion GmbH. Only selected parts of the program implemented by the student will be available in the IS, exclusively in the form of read-only source code under a proprietary license.