Learning lexical-syntactic biases: An fMRI study on how we connect words and syntactic information

Abstract

Language processing often involves learning new words and how they relate to each other. These relations are realized through syntactic information connected to a word, e.g. a word can be verb or a noun, or both, like the word 'run'. In a behavioral and an fMRI task we showed that words and their syntactic properties, i.e. lexical items which were either syntactically ambiguous or unambiguous, can be learned through the probabilities of co-occurrence in an exposure session and subsequently used in a production task. Novel words were processed within regions of the language network (left inferior frontal and posterior middle temporal gyrus) and more syntactic options led to higher activations herein, even when the words were shown in isolation, suggesting combined lexical-syntactic representation. When words were shown in untrained grammatical contexts, activation in left inferior frontal cortex increased. This might reflect competition between the newly learned representation and the presented information. The results elucidate the lexical nature of the neural representations of lexical-syntactic information within the language network and the specific role of the left inferior frontal cortex in unification of the novel words with the surrounding context.In order to produce and understand sentences, we must know the meanings and the grammatical rules of the language. Part of our grammatical knowledge is tightly linked to knowledge about individual words. We know, for instance, that the word 'cook' can be either a noun or a verb, but that the semantically similar word 'chef' can only be a noun. Similarly, 'to buy' and 'to acquire' have closely related meanings but only 'to buy' can be used both in a double object construction, as in 'The girl bought the boy a flower." and with a prepositional object, as in "The girl bought a flower for the boy.". Thus, we must have learned, through exposure to the languages, which types of syntactic options go with which words. This holds for all kinds of lexical-syntactic information, from word category to verb argument structure information. In this study we investigated how this type of lexically-bound syntactic information is learned, stored, and processed in the brain.Specifically, we looked at the differential contributions of core regions of the language network, left inferior frontal cortex and left middle temporal gyrus, to these processes.When processing words in context we activate a large left-dominant fronto-temporal network (Friederici and Gierhan 2013; Hagoort 2014; Hagoort and Indefrey 2014). Some models of language processing, such as the Memory, Unification and Control model (Hagoort 2005(Hagoort , 2013, make specific predictions about how lexical-syntactic information might be represented and processed within this language network. The assumption is that syntactic information, especially lexically-bound syntactic information, is stored with the lexical items in the 'mental lexicon' (Levelt 1992), located in posterior temporal regions of the brain. When words are...

A review of the neurobiology of syntax

“…: rules derived from real grammatical rules in natural language) compared to unreal ones (Musso et al, 2003) or to a phrase-structure grammar compared to a finite-state grammar (Friederici et al, 2006). It is also implicated in a study of word category learning (Weber et al, 2019). A similar pattern was found when contrasting the activity of a group of participants trained on the artificial grammar with an untrained group (Chen et al, 2019).…”

“…There is also evidence that the pMTG plays a crucial role in the storage of the lexicon Poeppel, 2004, 2007) which is consistent with observations from TMS (Krieger-Redwood and Jefferies, 2014) or aphasia (Dronkers et al, 2004) that lesions to the MTG impair lexical access. As such, there are proposals that the pMTG serves as the interface between the lexicon and syntax (Bozic et al, 2015;Weber et al, 2019;Caucheteux et al, 2021). In the model, the pMTG supports hierarchy building by both storing treelets in the lexicon, as well as combining or decomposing them.…”

“…where it is manipulated in the Unification space (Hagoort, 2013(Hagoort, , 2017Weber et al, 2019). While the multimodal nature of the IFG for unification does have some empirical support (Rodd et al, 2010;Kunert et al, 2015), this account would predict that damage to the IFG would have devastating effects for both productive and receptive syntax, as creating a full parse of a hierarchical structure would be impossible.…”