Slide Slam N11
Lexical neighbors support stronger decoding than non-word neighbors: Implications for the neural representation of words
Enes Avcu1, David Sorensen2, Seppo Ahlfors3, David Gow1,3,4; 1Massachusetts General Hospital / Harvard Medical School, 2Harvard University, 3Athinoula A. Martinos Center for Biomedical Imaging, 4Salem State University
Introduction: Neural decoding research has shown that cortical regions of the superior temporal gyrus are selectively sensitive to phonetic features. It is less clear how this segment-level input is bound into representations of wordform associated with downstream structures including the posterior middle temporal gyrus (pMTG) and temporal poles. Evidence for lexical neighborhood effects in spoken word recognition suggest that overlapping words partially activate multiple lexical candidates that actively compete. Such activation is consistent with either graded holistic or discrete compositional activation. Discriminating between these alternatives will provide insight into diverse phenomena including nonword lexical similarity effects, the perception of coarticulated speech, constraints on phonotactic structure, form priming and lexical competition effects. We hypothesize that classifiers trained to distinguish between lexical neighbors of two words (e.g. pig and toad) should also be able to distinguish those words without additional training. If lexical representation is holistic, non-words would not produce similar transfer learning. On the other hand, if lexical representation is compositional (e.g. segments or bigrams) training on nonword neighbors of the same words should also show similar transfer because those items will activate the same components as lexical neighbors do. Aim: To determine whether word level representation is holistic or compositional. Methods: The stimuli consisted of spoken CVC words and nonwords. Six hub words were chosen to define lexical neighborhoods. For each hub, we created 3 word and 3 nonword neighbors by changing only one phoneme, with the position of the changed phoneme counterbalanced across the 3 potential positions. For example, for the hub word pig we identified the lexical neighbors big, peg and pick and the nonword neighbors tig, poog and pid. Training and testing were done with 8 different talkers (4 male and 4 female). Simultaneous MEG/EEG data were collected (with structural MRI in a separate session) from 12 native English speakers while they complete an auditory lexical decision task. Three language-related ROIs (supramarginal gyrus, pMTG, temporal pole) and two control ROIs were created for each individual participant. Decoding was done based on each individual subject’s ROI source time courses using support vector machines (SVM). SVMs were trained to discriminate neighborhoods using trials when neighbors were presented then tested on untrained hub words to measure if transfer learning occurred. The pairwise classification analyses were repeated with different sets of neighbors. Separate analyses were done using only word or nonword neighbors, respectively, for training to examine the effect of lexicality. Results and Conclusion: SVMs successfully discriminated hub words after training on word neighborhoods with pMTG and temporal pole, but not with control region activation timeseries. However, training on nonword neighbors resulted in poorer classification. These results are consistent with the predictions of the holistic wordform representation hypothesis and demonstrate the potential utility of using neural decoding to explore the nature of lexical representation.