GIEO-Gujarati Information Extraction with Ontology-Driven BILSTM-CRF and INDICBERT Framework

Abstract

The rapid growth of unstructured Gujarati text in fields of news, government documents, and legal papers poses an enormous challenge for automated processing and information retrieval. The prevalent NLP resources for the Gujarati language are still limited, restraining the development of accurate language technologies. This effort proposes a deep learning-based information extraction framework that harnesses BiLSTM-CRF models with IndicBERT embeddings for constructing a full pipeline for POS tagging, chunking, and named entity recognition. The system uses resources like the Universal Dependencies Gujarati Treebank, IndicNER, and WikiAnn, apart from domain-custom annotated datasets. An ontology is then designed to provide a conceptual structure representing entities, relations, and attributes so that the extracted information can be stored as RDF triples and queried using SPARQL. Evaluations yield state-of-the-art results with an overall accuracy of 96.48%, including POS tagging and chunking at 96.10%, and NER F1- scores in excess of 90%. The ontology-based knowledge base supports various downstream applications such as question answering, decision support, and structured knowledge construction for Gujarati.