Presentation
Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Poster Slams
Bilateral temporal involvement in predictive morphological segmentation and processing during spoken word comprehension: MEG evidence from Arabic
Poster C40 in Poster Session C, Friday, October 7, 10:15 am - 12:00 pm EDT, Millennium Hall
Also presenting in Poster Slam C, Friday, October 7, 10:00 - 10:15 am EDT, Regency Ballroom
Suhail Matar1, Alec Marantz1,2; 1New York University, 2NYUAD Research Institute, NYU Abu Dhabi, UAE
A major challenge in speech processing is how the brain segments a continuous input stream. This has typically been studied on the phoneme or word level. But does the brain segment spoken words to their smallest meaningful units (i.e., morphemes)? If so, what’s the nature of this process? We contrast three hierarchically-nested models: (i) A simple morphologically-naïve model that has acoustic, lexical and phonetic predictors, but is insensitive to morphological information/boundaries; (ii) A morphologically-passive model, sensitive also to morpheme boundary/onset, and (iii) A morphologically-predictive model, sensitive also to predictive segmentation and morphological surprisal and uncertainty (calculated based on transition probabilities between morphemes). 27 participants listened to single words in Arabic while we recorded brain activity using magnetoencephalography (MEG). Words had two morphemes: a verb stem, and one of four direct object pronouns (e.g., ‘qayyama-ni’=‘(He) evaluated-me’; hyphens represent morpheme boundaries). Verb stems were either long (all following the Arabic template ‘_a__a_a’) or short (a shorter template with the same onset: ‘_a__a’). In Arabic, sets of root consonants are substituted into the underscored slots of these templates, producing different verbs (e.g., root {j,r,b} produces ‘jarraba’=‘(he) tested’ in the long template; {j,r} produces ‘jarra’=‘(he) dragged’ in the short template). We had two conditions. Morphologically unambiguous stems were all long (e.g., ‘qayyama’), with no derivable shorter stems (i.e., ‘qayya’ is not an Arabic verb/stem). Morphologically ambiguous stems were either short or long: all long stems had corresponding shorter stems with the same onset (e.g., ‘jarra’ vs. ‘jarraba’), producing temporary ambiguity during comprehension. All stems across conditions had identical uniqueness points (offset of the string _a__a). Comprehension tasks targeting either stems or pronouns followed 25% of trials. Using source-localization, we estimated cortical activation in bilateral inferior frontal, and superior/middle temporal cortex. We used a temporal response function (TRF) framework to estimate typical responses to each model’s predictors, measuring each model’s power to explain cortical activity. Compared to a null model, all three models significantly explained activity in all ROIs (p<0.0001;p-values corrected for multiple comparisons), which validates our models’ predictive power. When comparing nested model pairs, the passive model with morpheme onset information explained significantly more activity than the naïve model in bilateral temporal and inferior frontal ROIs (p=0.0001). In turn, the predictive model explained more activity than the passive model in bilateral superior temporal cortex (p<0.0001). This supports models of speech processing that contain bottom-up and top-down morphological information. We also compared evoked responses across conditions by averaging long-stem stimuli time-locked to uniqueness points (t=0). We found an early effect, 200–50ms before uniqueness point, in the bilateral superior temporal cortex, showing more negativity for ambiguous vs. unambiguous stems (p=0.002). This could index ‘eager’ predictive processing of a potential morpheme boundary in ambiguous stems. A later effect, 50–125ms after uniqueness point (left temporal:p=0.006; right temporal:p=0.02), shows the opposite pattern, possibly reflecting boundary revision in ambiguous stems. Our results provide evidence for morphological segmentation during speech processing, and support models where the brain predictively segments words into morphemes, rather than passively waiting for uniqueness points or boundaries.
Topic Areas: Morphology, Speech Perception