Analogies and semantic axes in large language models' embedding space
Anushree Divekar
Undergraduate Researcher
Marketing Major (Kelley School of Business)
YY Ahn
Faculty Mentor
YY Ahn (Luddy School of Informatics, Computing and Engineering)
Project Description
Deep learning (e.g., large language models like ChatGPT) works by learning highly useful vector-space embedding of entities such as words and images. It has been shown that this embedding space captures semantic relationships as geometric relationships (e.g., "king" - "man" + "woman" ~ "queen"). This research investigates the extent of these analogies and semantic axes in the embedding space of sentences and scientific articles.
Technology or Computational Component
This project will utilizes some large language models like sentence BERT and uses vector-based operations to understand the embedding space. Python will be used for the project.