Weakly-Supervised Learning Strategy for Domain-Specific Entity Recognition

The demand for accurate information extraction architectures for domain specific text data is steadily increasing. Annotated target domain datasets are necessary for finetuning Pretrained Language Models (PLMs) to perform on downstream tasks such as Named Entity Recognition (NER). The process of creating annotated datasets is often time-consuming and resource-intensive. Addressing the constraint imposed by the necessity for annotated datasets, we propose a new approach to weakly-supervised training pipelines, demonstrated through the use cases in automotive (Automotive Entity Recognition) and financial (Stock Entity Recognition) domains. The proposed approach achieves a remarkable F1 score of 0.855 on the financial dataset and 0.436 in the challenging domain of automotive texts. Our weakly-supervised training pipeline is characterized by a forty two-fold savings in financial resources and an excellent performance-to-cost ratio.