Macquarie University
Browse
01whole.pdf (4.42 MB)

Understanding predictive coding during language comprehension by comparing computational linguistic model with the MEG data from adult humans

Download (4.42 MB)
thesis
posted on 2022-08-03, 00:05 authored by Poonam Kumari

Language comprehension is an important task in the daily activities of human beings. It relies on the ability of human brain to correctly process the meanings of words and phrases, grammar, and text structure. It is believed that predictive coding is a fundamental mechanism of language comprehension in humans. Predictive coding is a theory of cognition in which the brain is constantly generating and updating a model of sensory input based on surprisal. Surprisal, a concept in psycholinguistics, refers to the degree of mismatch between what a human listener expected to be said next and what is really spoken. In this work, I aim to study prediction during language comprehension in humans by developing an n-gram based computational linguistic model of prediction from the text data of naturalistic speech stimuli, and test how well this model correlates with the language prediction related magnetoencephalography (MEG) signals of human brain. I used a MEG database of 16 adult humans, who were presented with audio stimuli (ABC podcasts). For the training of my n-gram model, I developed a domain-specific corpus using science and environment related articles from various web sources and online newspapers. I used the transcripts of my audio stimuli to test my n-gram model and calculated per-word surprisal. I then performed simple linear regression to find a relationship between per-word surprisal values and MEG data. Second level group statistics performed on the predictor parameter of our regression model showed significant correlation between surprisal and MEG signals between 140 and 320 milliseconds after word onset. These results are consistent with the contention that the brain responds to violation of predictions in a graded manner during the comprehension of speech stimuli. Future work will include the development of a larger domain-specific corpus and use computational linguistic models based on deep learning neural networks.

History

Table of Contents

Chapter 1: Introduction -- Chapter 2: Literature review -- Chapter 3: Methods -- Chapter 4: Results -- Chapter 5: Discussion -- Chapter 6: Future work -- References -- Appendix

Notes

Supervisor: Associate Professor Paul Sowman Submission date: 11th November 2019

Awarding Institution

Macquarie University

Degree Type

Thesis MRes

Degree

Thesis (MRes), Macquarie University, Faculty of Medicine, Health and Human Sciences, 2020

Department, Centre or School

Department of Cognitive Science

Year of Award

2020

Principal Supervisor

Paul Sowman

Rights

Copyright disclaimer: https://www.mq.edu.au/copyright-disclaimer Copyright Poonam Kumari 2020.

Language

English

Extent

43 pages

Usage metrics

    Macquarie University Theses

    Exports

    RefWorks
    BibTeX
    Ref. manager
    Endnote
    DataCite
    NLM
    DC