Gabriel Mora
Pronouns: He, Him, His
Research Mentor(s): Jonathan Brennan
Research Mentor School/College/Department: Linguistics / LSA
Program:
Authors:
Session: Session 3: 11:00 am – 11: 50 am
Poster: 45
Abstract
Despite the importance of the human brain and how it processes language, we know relatively little about how that processing works internally. One way to tackle this issue is to compare the properties of computational models for language from natural language processing (NLP) to signals recorded by brain measurements used in neurolinguistics such as electroencephalography (EEG). In this project, we analyze the similarities between EEG signals recorded from 33 participants, and 12 layers of GPT-2 embeddings, both from listening to one chapter of Alice in Wonderland. We developed a pipeline using Python MNE to enable this analysis. This pipeline begins by cleaning and preprocessing raw EEG signals and tokenizing them word by word. Then, we perform representational similarity analysis to create representational similarity matrices across the 2130 different words from the chapter using cosine distance to capture word to word correlation from the EEG data, while focusing on lexical/content words. We additionally compute the language comprehension activations in GPT-2 on the same word set, and build representational similarity matrices for each layer of GPT-2. We use these matrices to analyze similarities and differences both across subjects and between GPT-2 and human processing. This is done by comparing the time points at which EEG signals become most closely related to the GPT2 embeddings. This will show whether there are structures or patterns in human language processing that can be simulated and mathematically quantified by artificial intelligence and neural networks, which may help with some questions surrounding whether minds are entirely, partially, or not at all imitated.
