Poster B72, Tuesday, August 20, 2019, 3:15 – 5:00 pm, Restaurant Hall
Neural Computations of Prediction Error Can Explain MEG Responses during Recognition of Spoken Words and Pseudowords
Yingcan (Carol) Wang1, Ediz Sohoglu1, Rebecca Gilbert1, Richard N. Henson1, Matthew H. Davis1;1MRC Cognition and Brain Sciences Unit, University of Cambridge
INTRODUCTION. Listeners recognise familiar words spoken in their native language with a speed and accuracy that is unmatched by artificial systems. One impressive aspect of human perception is the ability to identify words while detecting and encoding unfamiliar pseudowords. One putative computational mechanism is Predictive Coding (PC) (Davis & Sohoglu, in press), by which computations of prediction error (the difference between heard and expected speech segments) support both the recognition of familiar words (captain) and detection of novel words (captick). METHODS. Twenty-four participants made lexical decisions on 160 sets of spoken words (e.g. captain, /kæptɪn/) and pseudowords (e.g. captick, /kæptɪk/) either with or without prior auditory presentation of another item sharing the same initial syllable (e.g. hearing captive, /kæptɪv/, or captiss, /kæptɪs/, before captain or captick). While performing this task, participants’ brain activity was recorded using concurrent electro/magnetoencephalography (EEG/MEG). Neural responses were time-locked to the deviation point (DP) where stimuli begin to acoustically diverge from each other (e.g. immediately after /ɪ/ for /kæptɪ…/). RESULTS. Behavioural results showed lexical competition effects (Monsell & Hirsh, 1998). Specifically, word recognition was significantly delayed by prior presentation (priming) from another word sharing the same initial sounds. However, behavioural effects of competitor priming were absent for trials with pseudoword primes or pseudoword targets. MEG results indicated that from around 400 to 700ms post DP, pseudowords elicited significantly stronger neural responses in the superior temporal gyrus (STG) than words. In the same time window and sensory locations, neural signals evoked by words and pseudowords showed an interaction between lexicality and priming in line with behavioural effects. Specifically, words (captain) primed by another similar sounding word (captive) evoked stronger neural responses in the STG than unprimed words. Whereas neural signals evoked by primed and unprimed pseudowords (captick) did not differ, regardless of the lexical status of the priming item (captain vs captiss). CONCLUSION. A number of different computational accounts can explain why STG responses elicited by pseudowords compared to words might differ (e.g. due to additional processing difficulty for segments in pseudowords, or greater segment prediction error). However, our observation that changes to the ease of word identification modulate the same STG responses appears more consistent with the PC model than other accounts (Davis & Sohoglu, in press). In the PC account, competitor priming is due to word identification strengthening predictions between pre- and post-DP segments of primed words (captain). Correspondingly larger PEs will be elicited by subsequent presentation of a lexical competitor (captive) time-locked to the divergent post-DP segments. In contrast, pseudoword prime presentations do not affect word identification or PE computations in the same way. Instead, pseudoword presentations elicit maximal PE signals that promote memory encoding of pseudowords but do not modify segment predictions. Overall, our study links neural computations of prediction error to lexical competition and pseudoword encoding processes, in line with PC principles.
Themes: Speech Perception, Perception: Speech Perception and Audiovisual Integration
Method: Electrophysiology (MEG/EEG/ECOG)