METAEXTRACT: A HYBRID METADATA EXTRACTION SYSTEM COMBINING RULE- BASED AND STATISTICAL APPROACHES

The issue of automatic extraction of structural metadata from documents is a widely studied topic in the field of document analysis, in which two main paradigms — universal models based on deep learning and rule-based systems — compete. Universal models have high generalization capabilities, but they require large amounts of annotated training data, which is a serious limitation for low-resource languages and domains. In this paper, a hybrid architecture combining a rule-based geometric-textual approach with a statistical named entity recognition (NER) model is presented through the example of extracting bibliographic metadata from scientific journal articles. The proposed system is based on the principle of creating rules from a single sample document and applying them to other documents sharing the same template, which makes it practically applicable even without large amounts of training data. The paper discusses design principles, the justification of architectural decisions, implementation details, and the role of this approach in the broader field of document extraction.

Hali tarjima qilinmagan