Clarissa Man
Research Mentor(s): Lisa Levinson
Mentor Department: Linguistics
Authors: Lisa Levinson, Clarissa Man
Session: Session 4 (1:00pm – 1:50pm)
Presentation Type: Poster 75
Abstract
A-maze is an improved version of the Maze task, a method used in psycholinguistics to study how people process language. In a Maze task, participants read a sentence one word at a time, choosing between two options at each step—one that correctly continues the sentence and another that does not. A-maze enhances this process by using natural language processing (NLP) to automatically generate more realistic incorrect word choices, making experiment design easier for researchers. It also enables efficient online data collection from a wide range of participants while accurately detecting moments of difficulty in language comprehension. This innovation overcomes key limitations of traditional methods like self-paced reading and eye tracking, providing a valuable tool for studying how people understand language. Building on the original A-maze, we have expanded its capabilities with multilingual support and improved accessibility. Using Hugging Face models like GPT-2, it now covers much more languages, with specialized tokenization for Mandarin. A new web-based interface, accessible via Google Colab, allows researchers to run experiments without local Python setup, simplifying data upload, parameter tuning, and result downloads. My work focuses on improving the algorithm that generates alternative words. Specifically, I have been developing a logging system that tracks how these alternatives are generated, ensuring transparency and clarity in the process.
