Skip to Main Content
 

Global Search Box

 
 
 
 

Files

File List


shain_thesis_2021-04-21.pdf (19.51 MB)

ETD Abstract Container

Abstract Header

Language, time, and the mind: Understanding human language processing using continuous-time deconvolutional regression

Shain, Cory Adam
http://rave.ohiolink.edu/etdc/view?acc_num=osu1619002281033782

Abstract Details

2021, Doctor of Philosophy, Ohio State University, Linguistics.
The predictions of theories of incremental human sentence processing are often cached out in word-by-word measures, but the mind is a dynamical system that responds to language in real time. As a result, there may be a complex alignment between the properties of words in language and the influence those properties exert on measures of human cognition. One possible aspect of this alignment is temporal diffusion, whereby sentence processing effects are realized in a delayed manner (Mitchell, 1984). For example, because of real-time bottlenecks in human information processing (Mollica & Piantadosi, 2017), encountering a surprising word may increase cognitive load not only at that word, but also on subsequent words as the rest of the experiment unfolds (Smith & Levy, 2013). In this thesis, I argue that effect timecourses are of direct or indirect importance to many central questions in psycholinguistics, that failure to account for these timecourses can have large impacts on the results of scientific hypothesis tests, and that existing discrete-time approaches to estimating and controlling for effect timecourses are not well adapted to many experimental designs in psycholinguistics, which involve non-uniform time series in which events (words) have variable duration. I define and implement an analysis technique that addresses these concerns: continuous-time deconvolutional regression (CDR). CDR estimates continuous-time functions that describe the shape and extent of a predictor's influence on the response over time, thus directly illuminating and controlling for temporally diffuse effects. I show empirically that CDR accurately recovers ground-truth models from synthetic data and provides plausible and detailed estimates of temporal structure in human data that generalize better than estimates obtained using existing techniques. I apply CDR to measures of naturalistic sentence processing in order test several theoretical questions in psycholinguistics. In one study, I present a CDR analysis that challenges an existing hypothesis that human reading times exhibit distinct effects of a word's overall frequency vs. its predictability from context (Staub, 2015), instead finding no evidence for such a distinction in naturalistic reading. In another study, I present a CDR analysis showing evidence from naturalistic functional magnetic resonance imaging (fMRI) data that human predictive mechanisms for language processing are sensitive to the syntactic features of sentences, and that these predictive mechanisms reside primarily in regions of the brain that are selective for language processing, rather than in regions involved in domain-general executive control. In a final study, I reanalyze the fMRI data with respect to theory-driven measures of working memory retrieval difficulty and report significant retrieval effects over strong controls for word predictability, but only in language-selective regions. This result supports a core role for language-specific working memory resources in typical language comprehension. Finally, I define and implement a deep neural generalization of CDR - the continuous-time deconvolutional regressive neural network (CDRNN) - that relaxes many of CDR's simplifying assumptions while retaining its deconvolutional interpretation. I show empirically that CDRNN generalizes better than CDR and other baselines on human data, and that it supports novel insights into the functional form of effects and effect interactions over time that are difficult to obtain using other methods. Based on these results, I advocate both increased attention to the temporal dimension in psycholinguistic regression analyses and the use of CDR to understand the dynamics of human sentence processing.
William Schuler, PhD (Advisor)
Micha Elsner, PhD (Advisor)
Paul Subhadeep, PhD (Committee Member)
283 p.
Computer Science; Linguistics; Neurosciences
deconvolution; impulse response; psycholinguistics; regression; temporal diffusion; time series; deep learning; prediction; integration; working memory; reading; fMRI; surprisal; dependency locality

Recommended Citations

Citations

  • Shain, C. A. (2021). Language, time, and the mind: Understanding human language processing using continuous-time deconvolutional regression [Doctoral dissertation, Ohio State University]. OhioLINK Electronic Theses and Dissertations Center. http://rave.ohiolink.edu/etdc/view?acc_num=osu1619002281033782

    APA Style (7th edition)

  • Shain, Cory. Language, time, and the mind: Understanding human language processing using continuous-time deconvolutional regression. 2021. Ohio State University, Doctoral dissertation. OhioLINK Electronic Theses and Dissertations Center, http://rave.ohiolink.edu/etdc/view?acc_num=osu1619002281033782.

    MLA Style (8th edition)

  • Shain, Cory. "Language, time, and the mind: Understanding human language processing using continuous-time deconvolutional regression." Doctoral dissertation, Ohio State University, 2021. http://rave.ohiolink.edu/etdc/view?acc_num=osu1619002281033782

    Chicago Manual of Style (17th edition)

osu1619002281033782
163
© 2021, some rights reserved.Creative Commons License Language, time, and the mind: Understanding human language processing using continuous-time deconvolutional regression by Cory Adam Shain is licensed under a Creative Commons Attribution-NonCommercial-NoDerivs 3.0 Unported License. Based on a work at etd.ohiolink.edu.
This open access ETD is published by The Ohio State University and OhioLINK.