Jim Xiong
Research Mentor(s): Alanson Sample
Mentor Department: Computer Science and Engineering
Authors: Jim Xiong, Yasha Iravantchi, Alanson Sample
Session: Session 5 (2:00pm – 2:50pm)
Presentation Type: Poster 13
Abstract
With the rapid pace of research today, keeping up with trends and understanding relationships between scientific papers is becoming more challenging than ever. Recognizing the growing difficulty researchers face in navigating an overwhelming volume of scientific literature, this project introduces a tool designed to address this problem by analyzing abstracts and titles using a large language model (LLM), a machine learning system trained on vast amounts of text to recognize patterns and extract meaning. The tool starts by creating a dataset using a web scraper to gather abstracts and titles from a variety of publications and conferences. A Mistral LLM processes this data, generating embeddings, or numerical representations that capture the meaning and context of each paper. To ensure high-quality embeddings, the Mistral model was fine-tuned and compared against more expensive models to achieve similar performance. The embeddings are then visualized using t-SNE, an algorithm that reduces a complex, high-dimensional dataset into a two-dimensional more easily visualizable format. This display highlights clusters of related topics, overlaps in themes, and outliers, giving users a clear and intuitive view of the research landscape. Beyond the visualizations, the tool lets users input any title or abstract to find similar papers, making it easy to explore related work and uncover connections. The tool assists researchers in spotting emerging trends, identifying knowledge gaps, and learning possible areas of interdisciplinary collaboration. Looking ahead, plans include speeding up data processing, improving the accuracy of the LLM’s embeddings, and enhancing the overall user experience. For future applications, this project could be used to make the research process more efficient and accessible.
