Farjana Sultana Samia

Research Overview

My research is in natural language processing, focused on connecting unstructured text to structured knowledge. I work on entity linking, information extraction and knowledge graphs, and on using large language models to reduce the cost of building labeled data for specialized scientific domains. This work is carried out in the Data and Knowledge Mining Lab at Iowa State University, advised by Prof. Dr. Qi Li.

Publications

• Bridging Lexical Divergence: LLM-Assisted, Cost-Efficient, Zero-shot Scientific Entity Linking

  PDF arXiv

  Md Rasel Khondokar, Qiao Qiao, Farjana Sultana Samia, Nhat Le, Yuepei Li, Qi Li

  Findings of EMNLP 2026
  • Introduces Sci-ZSEL, a zero-shot entity linking framework that generates entity aliases with large language models, filters noisy generations, and builds pseudo-labeled training data without any human annotation.
  • Targets the low lexical overlap between mentions and entity names in scientific text, where general-domain linking models transfer poorly.
  • Contributes a new animal science entity linking benchmark grounded in three livestock trait ontologies, with evaluation across five datasets.
  My contribution Built the pseudo-labeled data construction and filtering pipeline, which produces training data for fine-tuning without any human annotation.
  Topics

  • Entity linking
  • Zero-shot learning
  • Large language models
  • Scientific text
  • Ontologies
  • Pseudo-labeling

Current Research

• Natural Language Processing — Data and Knowledge Mining Lab, Iowa State University

  • Linking mentions in scientific and biomedical text to entities in structured knowledge bases and domain ontologies.
  • Relation extraction — identifying and classifying the relations that hold between entities mentioned in scientific literature.
  • Information extraction and knowledge graph construction from unstructured literature.
  • Zero-shot and low-supervision methods that use large language models in place of manual annotation.
  • Retrieval-augmented generation and retrieval over knowledge-intensive text.
  Topics

  • NLP
  • LLM
  • Entity linking
  • Relation extraction
  • Information extraction
  • Knowledge graphs
  • Retrieval-augmented generation

Previous Research

• Prediction of COVID-19 from Chest X-ray Images Using CNN and Pre-Trained Models

  Undergraduate thesis
  • Trained and compared seven pre-trained convolutional architectures on a COVID-19 chest X-ray dataset: VGG16, VGG19, ResNet50, InceptionV3, EfficientNetB0, NASNetMobile and DenseNet201.
  • VGG16 performed best at 97.67% accuracy; EfficientNetB0 performed worst at 50.00%.
  • On the VGG16 confusion matrix, 8 normal cases were misclassified as COVID-19 and 6 COVID-19 cases were misclassified as normal.
  Topics

  • Medical imaging
  • Transfer learning
  • CNN
  • Image classification

• Bangla Text Document Categorization

  • Automatic categorization of Bengali-language documents, motivated by the rapid growth of Bengali text on news portals, blogs and websites and the volume of unstructured data it produces.
  • Compared a range of machine learning and deep learning models for sorting and organizing large Bengali document collections.
  Topics

  • Bengali NLP
  • Document classification
  • Deep learning