Poster A26, Thursday, August 16, 10:15 am – 12:00 pm, Room 2000AB
Hierarchical syntactic structures modulate brain activity during morphological processing
Yohei Oseki1,2, Alec Marantz2;1Waseda University, 2New York University
A single-route "full decomposition" model of morphological processing (Taft, 1979, 2004; Taft & Forster, 1975) has proposed that there are three functionally different stages of morphological processing: morphological decomposition, lexical access, and morphological composition. In the recent literature on the neurobiology of language, these three stages have been associated with spatiotemporally dissociable evoked response components (Fruchter & Marantz, 2015): morphological decomposition around 170 ms in the inferior temporal/anterior fusiform cortex (i.e. M170), lexical access around 350 ms in the middle temporal cortex (i.e. M350), and morphological composition around 500 ms in the orbitofrontal cortex. However, given the theoretical agreement on hierarchical syntactic structures within words, what stage of morphological processing tracks hierarchical syntactic structures of words during morphological processing? In order to address this question, this paper conducts an magnetoencephalography (MEG) visual lexical decision experiment, where computational models proposed in natural language processing are employed to predict brain activity localized to Visual Word Form Area (VWFA), with special focus on the first stage of morphological decomposition indexed by the MEG evoked response component called the M170 (Gwilliams et al., 2016). The participants were 20 native English speakers. The stimuli were 800 novel morphologically complex trimorphemic words and nonwords with linear (root + suffix + suffix) and nested (prefix + root + suffix) syntactic structures. Two "amorphous" and three "morphous" computational models were investigated on the assumption that morphological processing proceeds incrementally from left to right. Specifically, the "amorphous" models, Letter Bigram Model and Syllable Bigram Model, estimate probabilities of morphologically complex words from bigram transition probabilities among letters and syllables, respectively, without reference to morphemes. The "morphous" models, Markov Model, Hidden Markov Model, and Probabilistic Context-Free Grammar, estimate probabilities of morphologically complex words from bigram transition probabilities among morphemes, linear syntactic structures, and hierarchical syntactic structures, respectively. The probability estimates were then transformed into surprisal (Hale, 2001; Levy, 2008), which was in turn employed as a predictor in regression analyses. Statistical analyses were based on two regions of interest (ROIs): anatomically-defined inferior temporal ROI and functionally-defined anterior fusiform ROI based on lemma frequency previously proposed to localize the M170 component. There are three primary results. First, all "morphous", but not "amorphous", models were statistically significant. Second, Probabilistic Context-Free Grammar, a computational model that estimates probabilities of morphologically complex words from hierarchical syntactic structures, most accurately predicted brain activity in the M170 evoked response component. Third, inspection of residual errors indicated that Probabilistic Context-Free Grammar explains nested words better than linear ones, while the opposite was the case for Markov Model and Hidden Markov Model. These results strongly suggest that the first stage of morphological decomposition indexed by the MEG evoked response component called the M170 not only reflects the decomposition of morphologically complex words into morphemes, but also the parsing of those complex words into hierarchical syntactic structures. Furthermore, morphological processing seems to track morphemes incrementally from left to right, contradicting "amorphous" models of morphological processing (Baayen et al., 2011). Consequently, morphological processing is sentence processing within words.
Topic Area: Grammar: Morphology