Neural bases of proactive and predictive processing of meaningful sub-word units in speech comprehension

Abstract

To comprehend speech, human brains identify meaningful units in the speech stream. But whereas the English She believed him. has 3 words, the Arabic equivalent saddaqathu. is a single word with 3 meaningful sub-word units, called morphemes: a verb stem ( saddaqa), a subject suffix ( -t-), and a direct object pronoun ( -hu). It remains unclear whether and how the brain processes morphemes, above and beyond other language units, during speech comprehension. Here, we propose and test hierarchically-nested encoding models of speech comprehension: a NAIVE model with word-, syllable-, and sound-level information; a BOTTOM-UP model with additional morpheme boundary information; and PREDICTIVE models that process morphemes before these boundaries. We recorded magnetoencephalography (MEG) data as participants listened to Arabic sentences like saddaqathu.. A temporal response function (TRF) analysis revealed that in temporal and left inferior frontal regions PREDICTIVE models outperform the BOTTOM-UP model, which outperforms the NAIVE model. Moreover, verb stems were either length-AMBIGUOUS (e.g., saddaqa could initially be mistaken for the shorter stem sadda= blocked) or length-UNAMBIGUOUS (e.g., qayyama= evaluated cannot be mistaken for a shorter stem), but shared a uniqueness point, at which stem identity is fully disambiguated. Evoked analyses revealed differences between conditions before the uniqueness point, suggesting that, rather than await disambiguation, the brain employs PROACTIVE PREDICTIVE strategies, processing the accumulated input as soon as any possible stem is identifiable, even if not unique. These findings highlight the role of morpheme processing in speech comprehension, and the importance of including morpheme-level information in neural and computational models of speech comprehension. Significance statementMany leading models of speech comprehension include information about words, syllables and sounds. But languages vary considerably in the amount of meaning packed into word units. This work proposes speech comprehension models with information about meaningful sub-word units, called morphemes (e.g., bake- and -ing in baking), and shows that they explain significantly more neural activity than models without morpheme information. We also show how the brain predictively processes morphemic information. These findings highlight the role of morphemes in speech comprehension and emphasize the contributions of morpheme-level information-theoretic metrics, like surprisal and entropy. Our models can be used to update current neural, cognitive, and computational models of speech comprehension, and constitute a step towards refining those models for naturalistic, connected speech.