Active Learning for Document Analysis - A Conceptual Framework of AL Scenarios

Even after having a huge amount of unlabelled, free data which is easy to collect, the process of labeling data is still time-consuming and costly, making it important to develop supervised learning models. With Active Learning (AL), it is possible to save the cost of labeling data. It makes querying of high-quality, limited data points easier by searching for the most representative and informative points. It increases generalizability and classification performance. This study introduces theoretical concept of active learning scenarios, techniques, and components and gives illustrative examples of how AL works and its benefits. This review paper provides strategies for AL scenarios and gives high-level summary to describe a lot of practical challenges with active learning. Finally, this study has also discussed the most widely-used active learning packages and evaluation metrics.