lu.se

Neurolingvistik

Humanistiska och teologiska fakulteterna | Lunds universitet

Neurolingvistik

The Lund Neurolinguistics group

The Lund Neurolinguistics group utilizes cutting-edge brain imaging techniques to assess the neural underpinnings of language processing and learning. Using the event-related potential (ERP) technique, we follow speech processing as it develops in real-time with millisecond precision. With 3 and 7 tesla functional and structural magnetic resonance imaging (fMRI/MRI), we observe regions active during different language tasks and measure the correlation between brain structure and language aptitude with high spatial resolution. The following sections present some major research areas and findings:

Pre-activation negativity (PrAN)
Left-edge boundary tones
Word accents
Dual route processing, word accents, and cortical thickness
Learning word accents and morphology
A thicker Broca's area is better for grammar learning
Time constraints on working memory and language processing

Pre-activation negativity (PrAN)

The pre-activation negativity (PrAN) is an electrically negative brain potential discovered in our lab. It shows the predictive strength of phonological cues. Already at word onset, PrAN has higher amplitude for word beginnings that have few possible word continuations of high lexical frequency (Söderström, Horne, Frid & Roll, 2016; Roll, Söderström, Frid, Mannfolk, & Horne, 2017).

PrAN and cohort size

The number of possible word continuations is also referred to as the word beginning's cohort size (Marslen-Wilson, 1984). Thus, word beginnings with smaller cohort sizes have a larger PrAN (Söderström et al., 2016). PrAN amplitude is also larger if the words of the cohort have high lexical frequency (Roll et al., 2017). Since PrAN is a negative component, its decrease with cohort size is positive and its increase with lexical frequency is negative:

PrAN = k (cohort size) – c (lexical frequency)

where k and c are constants.

Left-edge boundary tones

We have discovered a "left-edge boundary tone" (LEBT) occurring in the first prosodic word of Swedish main clauses but not in subordinate clauses (Roll, 2006 (PDF); cf. Myrberg, 2010 (PDF)). The complementizer att 'that' can introduce both main clauses and subordinate clauses, which would be temporarily ambiguous were it not for the LEBT. The presence or absence of a LEBT can therefore lead to garden path effects. For example, with no LEBT, listeners expect subordinate clause structure. If they instead hear a word order typical of main clauses, they have been lead down the garden path. Using ERP, reanalysis of the unexpected structure is reflected in a P600 effect (Roll, Horne, & Lindgren, 2009; 2011; Roll & Horne, 2011). Upon hearing that a LEBT is missing at the beginning of a clause, listeners strongly activate subordinate clause structure. This gives rise to a PrAN with its most likely sources in Broca's area (Söderström, Horne, Mannfolk, van Westen, & Roll, 2018).

Word accents

Swedish and Norwegian "word accents" consist of a high or a low tone on a word stem. The stem tone, however, is induced mainly by the word's suffix (Riad, 2014; Rischel, 1963). Thus, the stem hatt 'hat' is pronounced with a low tone (accent 1) in hatt+en 'hat+the,' but the plural suffix -ar induces a high tone (accent 2) onto the stem in hatt+ar 'hat+s.' The same melody change is found in all words with this variation: båten/båtar 'the boat/boats,' bussen/bussar 'the bus/buses' etc. Swedish word accents are often assumed to have low "functional load," since the contrast between word accents is hardly ever used to differentiate between words. However, we have found that, just like left-edge boundary tones that lead listeners to expect a certain sentence structure, word accents lead listeners to expect a particular word structure. Thus, a suffix that has been invalidly cued by the wrong word accent produces a P600, showing that the word structure needs to be reanalyzed (Roll, Horne, & Lindgren, 2010; Roll, Söderström, & Horne, 2013; Roll, 2015; Roll, Söderström, Shtyrov, Mannfolk, Johansson, van Westen, & Horne, 2015; Söderström, Horne, & Roll, 2017; Söderström, Horne, Mannfolk, van Westen, & Roll, 2018). Word accents even produce a pre-activation negativity (PrAN), showing how they pre-activate the ending of words (Roll, 2015; Söderström et al., 2016), possibly originating in Wernicke's and Broca's areas (Roll, et al., 2015; Söderström et al., 2018). We suggest that the main function of word accents might be the predictive rather than the distinctive function. Word accents make processing of their associated suffixes quicker (Söderström, Roll, & Horne, 2012), and a thicker cortex in Wernicke's area speeds up processing of the connection between word accent and suffix (Schremm, Novén, Horne, Söderström, van Westen, & Roll, 2018).

Dual route processing, word accents, and cortical thickness

Two different routes of processing are assumed to be available for morphologically complex words (Pinker, 1991). Common, frequent forms of words are thought to be stored as whole word forms for quick access, involving Wernicke's area and adjacent regions in the left temporal lobe. Infrequent or previously not encountered forms, on the other hand, can be decomposed into stem and affixes to access the stem meaning. In other words, we can understand an incorrect form like *springde *'runned' by decomposing the word into the stem stem and spring+de 'run+PST.' Decomposition, however, involves the 'dorsal route' for grammatical processing (Friederici, 2017), in particular the posterior portion of Broca's area. This processing is likely to be what is reflected in the ERP effect for misapplication of a regular grammatical suffix on an irregular stem, as in *spring+de *'runn+ed': a left anterior negativity (LAN) (Schremm, Novén, Horne, & Roll, 2019). Suffixes that have been incorrectly cued by the wrong word accent in pseudowords with real suffixes like kvup+ar 'kvup+PL' produce a LAN, indicating that an abstract rule associating tone and suffix might be at play (Söderström, Horne, & Roll, 2017). In fact, just like stem-suffix processing in general, tone-suffix association seems to be represented in two possible routes in the brain. Evidence for dual route representation of word accents comes from correlations between cortical thickness and Swedish speakers' capacity for processing their own native language. Thus, native speakers with thicker cortex in Wernicke's area process the word accent-suffix associations quicker in frequent real words than speakers with thinner cortex, suggesting that strong whole form representations are important for real words. Conversely, native speakers with thicker cortex in the posterior portion of Broca's area are quicker to process word accent-suffix association as an abstract rule in pseudowords with real stems (Schremm, Novén, Horne, Söderström, van Westen, & Roll, 2018).

Learning word accents and morphology

Since word accents are powerful predictors for the endings of words, they can be useful also for second language learners of Swedish to increment their speed and efficacy of word processing. Indeed, after several years of exposure, second language speakers seem to acquire the connection between word accent and suffix through implicit learning (Schremm, Söderström, Horne, & Roll, 2016). However, to accelerate the process, we designed a computer game for learning and automatizing the connection between word accent and suffix in a playful manner (Schremm, Hed, Horne, & Roll, 2017). The player navigates a dinosaur and jumps on platforms to make decisions as to which ending a word should have based on the tone on the word's stem. Beginner learners do not use word accents predictively (Gosselke Berthelsen, Horne, Brännström, Shtyrov, & Roll, 2018). Using a prototype of the game for two weeks, learners developed a pre-activation negativity for word accents (Hed, Schremm, Horne, & Roll, submitted). A version for learning Swedish morphology has also been developed.

A thicker Broca's area is better for grammar learning

In a spatially unbiased whole-brain analysis, we found correlation between cortical thickness in the frontal part of Broca's area and aptitude for grammar learning in adults (Novén, Schremm, Nilsson, Horne, & Roll, 2019). Specifically, participants with thicker cortex in this area scored higher on the LLAMA F (Meara, 2005) artificial grammar learning test. In the homologue area in the right hemisphere, participants with thinner cortex were better at pitch discrimination, in line with previous results (Hyde et al., 2007). It might be the case that a thinner, more streamlined cortex is better for tasks involving lower complexity, whereas high-complexity tasks benefit from a thicker cortex with increased possibilities of neural connections (Novén, Schremm, Nilsson, Horne, & Roll, 2019).

Time constraints on working memory and language processing

The form of words and sentences starts fading away in short-term memory after only 2 to 3 seconds (Peterson & Peterson, 1959; Baddeley et al., 1975). Conversely, the meaning of sentences still has a rather detailed representation in working memory after as long as 46 s (Sachs, 1967; 1974). The rapid decay of the form of words seems to have consequences on processing. Thus, using ERP, we found that prosodic breaks were spontaneously generated every 2.7 s during silent reading, regardless of whether one, two, or three clauses were read within this time frame (Roll, Lindgren, Alter, & Horne, 2012). Thus, by varying word presentation rates, we controlled the number of clauses that participants read in 2.7, and invariably observed a 'closure positive shift' (CPS) (Steinhauer, Alter, & Friederici, 1999) at this time constant. This implicit prosodic closure was also shown to induce early syntactic closure. A sentence adverb indicating late closure after a time constant-induced prosodic boundary produced an early negativity indicating error detection and a late positivity (P600) suggesting reprocessing (Schremm, Horne, & Roll, 2015). Grammatical relations also seem to be differently processed when the form of words is still present in working memory (< 2.7 s) and when only the meaning is left (> 2.7 s). In favor of this hypothesis, readers take longer to detect disagreeing words (e.g. *they runs) when they are separated by a time interval of over 2.7 s in object relative clauses (Schremm, Horne, & Roll, 2015b). In the ERPs, disagreement at distances under 2.7 s gives rise to a left-anterior negativity (LAN), whereas the negativity is rather slightly right-lateralized when the disagreeing words are separated by an interval exceeding 2.7 s (Roll, Gosselke, Lindgren, & Horne, 2013).