Automatically Domain-Adapted and Personalized Document Analysis

Document Analysis (DA) technologies are becoming increasingly pervasive in our daily lives due to the digitalization of documents (both in the cultural and industrial domains) and the widespread use of paper tablets, pads, and smartphones to take notes and sign documents. High-performing DA algorithms are needed that are able to deal with digitized documents from different writers, in different languages, and with different visual characteristics.

As a result, modern DA systems must be able not only to generalize, but also to adapt efficiently to new domains and users, often under limited supervision or data availability. In this context, domain adaptation and automatic personalization play a central role in improving the robustness and usability of DA techniques.

Recent advances in large pre-trained language and multimodal models enable new forms of lightweight domain adaptation of generic systems. Strategies such as prompt engineering, few-shot and in-context learning, structured output generation, and task-specific post-processing allow adapting these models to specific domains without full retraining.

Furthermore, in contexts involving sensitive information, privacy-preserving solutions and lightweight adaptation strategies that can be performed onboard personal devices are crucial.