More of that, please: Domain Adaptation of Information Extraction through Examples & Feedback

Automatic information extraction, e.g., into a tabular format, is crucial for leveraging knowledge in large text collections. Yet, creating such extraction pipelines for custom target attributes can cause high overheads, while off-the-shelf tools might miss domain-specific information. Therefore, in this paper, we propose an interactive system that augments generic extractions and aligns them with a target definition. The necessary domain adaptation is reached through examples provided by the users during the interaction with the system. As part of this, we propose different low-overhead extractors and evaluate them individually and end-to-end to demonstrate how our approach minimizes the necessary interactions. We publish our code as open source.