Project Study: Research on Name Matching Algorithms / Deutsche Bank AG

Banks manage enormous amounts of data, a significant portion of which is unstructured. Names of clients, companies and counterparties appear across many systems, often in different spellings and formats.

Today, when searching for information in Deutsche Bank’s data lake, the bank largely depends on exact matches and traditional full-text search techniques to identify relevant entries. However, this approach often struggles to uncover hidden relationships, context and semantically similar information, which limits the full value that can be extracted from the bank’s data assets. This project study therefore conducts research on how to intelligently compute name matching and string similarity.

The goal is to research and prototype approaches that intelligently compute name matching and string similarity, going beyond exact matches and traditional full-text search. The students:

(a) review the state of the art in name matching and string similarity (e.g., edit-distance measures such as Levenshtein or Jaro-Winkler, phonetic, token-and n-gram-based methods, and embedding-based semantic similarity);
(b) analyze typical challenges of names in banking data, such as spelling variants, abbreviations, transliterations, legal-form suffixes, word order and typos;
(c) develop a prototype that combines suitable methods and uses probabilistic and numerical techniques to compute meaningful similarity and match scores; and
(d) evaluate the approaches on suitable test data (e.g., accuracy, precision/recall, runtime) and derive recommendations for their use in the bank’s data landscape.