How to Improve E-commerce Search Engines? Evaluating Transformer-Based Named Entity Recognition on German Product Datasets

The quality of e-commerce search engines often suffers from data that online retailers poorly maintain. This situation can be observed on consumer-to-consumer marketplaces as well as on business-to-consumer platforms. One way to improve search quality is to perform linguistic enhancement of the product data. In this case, Named Entity Recognition is primarily used to identify important content and give it a higher weighting in the search. Our approach detects e-commerce entity types, such as products, brands, and various product attributes. Because of the low availability of existing resources and linguistic complexity identifying these entity types is challenging. Therefore, we acquire data from two online e-commerce marketplaces to build six German datasets based on product titles and descriptions. For these datasets, we evaluate the NER performance of the state-of-the-art models BERT, RoBERTa, and XLM-RoBERTa. The best performance archived the XLM-RoBERTa model with an F1 score of 0.8611 averaged over all datasets.