Poster A80, Tuesday, August 20, 2019, 10:15 am – 12:00 pm, Restaurant Hall
Recognizing words by their neighbors: Neural decoding evidence for overlapping lexical representations
Seppo Ahlfors1, Adriana Schoenhaut1, David Gow1;1Massachusetts General Hospital
Introduction: While a growing body of research has examined the neural representation of phonetic features, very little work has examined the neural representation of phonological wordforms. One fundamental question is whether lexical representations are compositional or holistic. If lexical representations are compositional (e.g. made up of subunits such as segments, syllables or N-phones), then words with overlapping phonological patterns may have overlapping neural representations. However, if lexical representation is holistic, phonologically similar words may not have similar neural representations. To examine these predictions, we trained classifiers to discriminate between sets of words and nonwords from different phonological neighborhoods, and tested their ability to discriminate between untrained words that define those neighborhoods using spatiotemporal activation patterns from brain regions associated with phonological wordform representation. Methods: MEG/EEG data were collected simultaneously while subjects performed an auditory lexical decision task. The stimuli included monomorphemic CVC seed words, and a combination of real words and nonwords derived by changing a single feature of one phoneme of those seed words (e.g. pig -> big, peg, pick, tig, poog, pid). Stimuli were recorded by male and female talkers, and those stimuli were digitally manipulated to create tokens of all stimuli in 8 different voices. Trials were blocked by voice, with all seed words heard twice and phonological neighbors heard once per block, and all voices presented in two separate blocks. Results: Subjects were able to perform the task with high accuracy. MRI-constrained minimum norm source estimates of MEG/EEG data were created for each trial, and automatically parcellated into ROIs including regions associated with wordform representation (supramarginal gyrus, posterior middle temporal gyrus – see Gow, 2012). We trained a support vector machine to discriminate between trials based on neighborhood membership (e.g. neighbors of pig versus neighbors of bike), and then used those classifiers to discriminate between neural responses to tokens of seed words (e.g. pig versus bike). Analyses showed strong transfer of neighbor training to seed discrimination. Moreover, transfer did not depend on the lexical status of the neighbors used to train the classifiers. Conclusion: The results show that activation patterns associated with phonological neighbors of seed words are sufficiently similar to those produced by seed words to support classification. This is consistent with compositional representation of wordform, with overlapping elements of seed and neighbor words evoking common patterns of activation. An alternative interpretation involving obligatory parallel activation of lexical wordforms that overlap with heard forms is also considered. Implications of these results for understanding both behavioral neighborhood effects and nonword wordlikeness effects are discussed.
Themes: Phonology and Phonological Working Memory, Perception: Speech Perception and Audiovisual Integration
Method: Electrophysiology (MEG/EEG/ECOG)