Within domain-specific data science, people often struggled with starting projects without decent training data. Especially for token-classification this is a time-consuming issue. Nowadays, there are many shortcuts that one can use to get a head start with projects like these by applying techniques like active learning, weak-supervision, few-shot learning, and cross-lingual models, however, in practice, people do not commonly use these.
The presentation shows how anyone can bootstrap a scarce data project by going over medical transcriptions. These transcriptions are pre-annotated by using a list of available keywords. These keywords are then evaluated using data exploration within the Argilla app. We further fine-tune our approach by applying few-shot NER with my concise-concepts package. Additionally, a weak-supervision rule base will be composed, which is actively being maintained using an active-learning listener as a background process during the annotation. Lastly, we combine all of these shortcuts into a SparkNLP training pipeline for NER.