Sankalp Sanjay Bahad supervised by Dr. Parameswari Krishnamurthy received  her Master of Science – Dual Degree in Computational Linguistics (CLD). Here’s a summary of her research work on From Named Entities to Knowledge Graphs: Information Extraction in Low-Resource and Biomedical Domains

Named Entity Recognition (NER) is a foundational task in Natural Language Processing (NLP) that supports applications such as relation extraction, question answering, and knowledge graph construction. Although transformer-based models achieve strong performance on English benchmarks, Named Entity Recognition (NER) remains challenging in morphologically rich low-resource languages and specialised domains such as biomedicine, where annotated data and unified schemas are limited. This thesis develops information extraction pipelines for four Indian languages and the biomedical domain under low-supervision settings. For Indian languages, we construct a human-annotated NER corpus of approximately 40,000 sentences spanning Hindi, Urdu, Odia, and Telugu under a six-type schema. A multilingual model fine-tuned on this corpus achieves an average F1 of approximately 0.80 and remains competitive on unseen Indic benchmarks. Using these outputs, we build a cross-lingual knowledge graph pipeline combining rule-based and zero-shot NLI relation extraction with LaBSEbased entity linking, extracting 15,126 triples with up to 80.75 For biomedicine, we unify multiple corpora with disjoint label inventories under a shared seventype schema and extend supervision through agreement-based pseudo-labeling, retaining mentions only when two independently trained teacher models agree on span and type. This improves entity-level F1 from approximately 0.55 to 0.69, with gains of up to 3.7 points on low-frequency entity types. The improved NER system is integrated into a complete Information Extraction (IE) pipeline evaluated across retrieval, synthetic graph Question Answering (QA), and PubMedQA tasks. Results show that NER quality is the primary bottleneck for downstream retrieval, while intrinsic evaluation substantially overestimates factual accuracy compared to evaluation against curated canonical facts. A companion study further extends the pipeline to document-level coreference resolution across five Indian languages. Overall, the thesis demonstrates that careful schema design, agreement-filtered pseudo-labeling, and lightweight cross-lingual methods can produce robust NER and IE systems in low-resource and specialised domains. 

June 2026