Named Entity Recognition on Transaction Descriptions

Abstract: With the surge of open banking, there is a large increase in applications based on transaction data. Therefore, there is a need for being able to extract important information from Swedish transaction descriptions in a structured way. We designed models for named entity recognition on transaction descriptions that can identify and classify organizations, locations, persons, payment providers (e.g Swish, Klarna, or PayPal) and products/apps. With our best NER model, we reached a chunk F1 score 0.849, despite only using 2200 transactions for training and transaction descriptions being messy. This is, to the best of our knowledge, the first published report on named entity recognition for transaction descriptions. Finally, we used our named entity recognition model and its output to complement a commercial transaction categorization system and improve the performance of the two existing models by 7.4% and 1.1% respectively.