Can I have a quick word? Early electrophysiological manifestations of psycholinguistic processes revealed by event-related regression analysis of the EEG

Abstract

We applied multiple linear regression analysis to event-related electrophysiological responses to words and pseudowords in a visual lexical decision task, yielding event-related regression coefficients (ERRCs) instead of the traditional event-related potential (ERP) measure. Our main goal was to disentangle the earliest ERP effects of the length of letter strings (“word length”) and orthographic neighbourhood size (Coltheart's “N”). With respect to N, existing evidence is still ambiguous with respect to whether effects of N reflect early access to lexico-semantic information, or whether they occur at later decision or verification stages. In the present study, we found distinct neurophysiological manifestations of both N and word length around 100 ms after word onset. Importantly, the effect of N distinguished between words and pseudowords, while the effect of word length did not. Minimum norm source estimation revealed the most dominant sources for word length in bilateral posterior brain areas for both words and pseudowords. For N, these sources were more left-lateralised and consistent with perisylvian brain areas, with activation peaks in temporal areas being more anterior for words compared to pseudowords. Our results support evidence for an effect of N at early and elementary stages of word recognition. We discuss the implications of these results for the time line of word recognition processes, and emphasise the value of ERRCs in combination with source analysis in psycholinguistic and cognitive brain research.

Introduction

The neuroscientific investigation of written word recognition faces the major problem that the processes of interest are affected by a large number of intercorrelated variables. The situation is complicated by the fact that highly correlated variables can affect different aspects of processing. For example, word length and orthographic neighbourhood size (i.e. the number N of words that can be obtained from a base word by exchanging just one letter, such as “can” into “car”), are negatively correlated in normal language. This reflects the fact that short words (such as “cat”) are commonly similar to more other words (such as “mat”, “fat”, “can”) than long ones (such as “crocodile”). While word length is commonly linked to early visual or orthographic processes (e.g. Ellis, 2004, Hauk and Pulvermüller, 2004a, Mechelli et al., 2000), effects of N have been interpreted in terms of competition during lexical access (Andrews, 1997, Grainger and Jacobs, 1996, Holcomb et al., 2002), or even post-lexical processing (Fiebach et al., 2007). Thus new methods of disentangling these correlated variables are valuable if neuroscientists are to discover correlates of these psycholinguistic variables in early or late processes during the recognition of visually presented words.

Most electrophysiological and neuroimaging studies so far have been using factorial contrasts (e.g. words versus pseudowords, long words versus short words, etc.) in order to determine the effect of a particular variable on the brain response. This approach has two main disadvantages: (1) it does not exploit information in the continuous distribution of values, e.g. of word lengths for individual items. (2) In order to match for highly correlated confounding variables that are not of interest (e.g. for word length when effects of N are studied), “unusual” items on the extremes of the parameter distributions might have to be chosen (Baayen et al., 1997, Ford et al., 2003).

An alternative to factorial designs is multiple linear regression analysis. This method allows testing to what degree a variable (e.g. word length) predicts data across all trials. In the case of only one variable, this corresponds to computing the covariance between the data with the predictor variable. In the case of multiple intercorrelated variables, orthogonalisation procedures need to be applied in order to obtain independent estimates for each variable, as will be described in more detail below. All other things being equal, regression designs will have greater power than dichotomizing continuous variables in a factorial experiment (Cohen, 1983).

Multiple linear regression analysis has long been applied to behavioural data on visual word recognition (e.g. Balota et al., 2004, Whaley, 1978). In neuroimaging, regression analysis is part of the commonly used “general linear model” (Friston et al., 1995), but it has not been widely applied in language research (see Davis et al., 2004, Graves et al., 2007, for examples). It has only very recently been introduced to the field of human electrophysiology (Dambacher et al., 2006, Hauk et al., 2006a). In the following, we will briefly summarise how electrophysiological studies on psycholinguistic variables have so far contributed to our knowledge about the early time course of visual word recognition. We will then motivate our interest in the particular variable N, which has been investigated by only few neuroscientific studies. We will also describe the multiple linear regression approach applied in the present study in more detail, emphasising aspects that are special to the analysis of electroencephalographic (EEG) or magnetoencephalographic (MEG) data.

So far, the most commonly assessed psycholinguistic properties are word frequency (estimate of the frequency of a word's occurrence in the language) and lexicality (difference between words and non- or pseudowords); two variables which also produce reliable effects in behavioural data (Gernsbacher, 1984, Whaley, 1978). In electrophysiological research, Sereno et al. (1998) reported effects of word frequency on the “N1” component at 132 ms in a lexical decision experiment. A recent ERP study using lexical decision found word frequency effects interacting with emotional quality of words even earlier, around 100 ms (Scott et al., 2008). Effects of word frequency were detected slightly later by the studies of Hauk and Pulvermüller (2004a) around 160 ms in a lexical decision task, and Dambacher et al. (2006) around 170 ms for sentence reading. Assadollahi and Pulvermüller (2003) found an effect of word frequency in their MEG on word reading around 150 ms, that surfaced as an interaction with the variable word length. An interaction between word length and frequency was also found for words presented in sentence context around 120 and 180 ms (Penolazzi et al., 2007). An even earlier effect of word frequency occurred in the EEG study of Hauk et al. (2006b) around 110 ms, employing a lexical decision task. In all of these studies, higher frequency words generally evoked lower amplitude neural responses.

Lexicality effects have been consistently reported around 200–250 ms, for example as a “word recognition potential” (Hinojosa et al., 2004, Martin-Loeches et al., 1999, Rudell, 1991). Similarly, Dehaene (1995) reported differences between words and consonant strings at 192 ms. Cohen et al. (2000) described ERP differences between words and non-words shortly after 200 ms. Hauk et al. (2006b) reported a main effect of lexicality around 200 ms, but an earlier effect of lexicality that interacted with orthographic typicality around 160 ms. Hauk et al. (2006a) found a significant difference between words and pseudowords around 160 ms. The earliest lexicality effects so far have been reported by Sereno et al. (1998) at 100 ms.

This overview demonstrates that effects of lexicality and word frequency before 200 ms have occurred across several different studies. These results have been corroborated and extended by a few studies using linear regression or related methods. In the study of Hauk et al. (2006b), words and pseudowords were presented during an EEG experiment in a lexical decision task. The effects of several psycholinguistic variables on brain responses were evaluated using multiple linear regression on the EEG responses to words to derive event-related regression coefficients (ERRCs). These ERRCs quantify the influence of specific parameters of written words on evoked electrophysiological responses (the regression equivalent of the “event-related potentials”, ERPs, derived from a traditional factorial design). Variables associated with the form of written words (word length and orthographic typicality as quantified by bi- and trigram frequencies) affected the ERRCs already around 90 ms after word onset. Effects of word frequency were detected shortly afterwards around 110 ms. The morpho-semantic variable Semantic Coherence (describing the consistency of meanings within a morphological family; Ford et al., 2003, Landauer and Dumais, 1997) was reflected in the ERRCs around 160 ms, co-occurring with the first significant difference between words and pseudowords.

This pattern of results was interpreted in terms of a serial but cascaded sequence of processing steps within the first 200 ms after presentation of a written word. This interpretation was further supported by source estimation results. Length produced right-lateralised activation in posterior brain areas consistent with basic visual analysis of the word form. Activation for orthographic typicality was left-lateralised, and the largest centre of activation occurred at a left inferior temporal location. The earliest influence of word frequency was also left-lateralised, and located in a posterior temporal area. The source estimates reflecting Semantic Coherence, as well as for the word/pseudoword differences, were more distributed across both hemispheres. Interestingly, around 200 ms as well as at later latencies between 300 and 500 ms, several psycholinguistic variables produced effects simultaneously and with similar topographies. This lack of specificity may indicate either integration of information across psycholinguistic processing levels, or post-access verification and decision processes.

Similar evidence for early lexical access for written words have also been obtained for sentence stimuli, Dambacher et al. (2006) presented sentences word-by-word to their subjects, and varied the lexical frequency of open-class words at different positions within the sentence, as well as their predictability by the context. Regression analysis revealed a context-independent effect (i.e. not depending on word position within a sentence) of word frequency around 170 ms after word presentation. The N400 component was affected by both predictability and word frequency. They suggested that lexical information is accessed before 200 ms, and that later ERP components reflect contextual integration processes. Dien et al. (2003) used items-analysis in their analysis of EEG responses to congruous or incongruous sentence endings, respectively. They averaged epochs for individual stimulus items across subjects, and applied parametric analysis to the resulting data set. The variables of interest were ‘meaningfulness’ and ‘expectedness’ of the target word (i.e. how much sense the sentence makes including the target words, versus how strongly the participants expect this word given the preceding context). Earliest effects of both variables were detected around 200 ms after word onset, suggesting that lexical information about both the target word and the context in which it occurs is already available at this latency. This is consistent with findings that effects of word class (e.g. verbs versus nouns) or semantic and emotional attributes (for example effector type of actions) have been reported around 200–250 ms (Hauk and Pulvermüller, 2004b, Kissler et al., 2008, Pulvermüller et al., 1995, Pulvermüller et al., 1999, Skrandies, 1998).

Another variable that potentially taps into lexical processing, but has only been investigated in a few electrophysiological and neuroimaging studies, is orthographic neighbourhood size, or N (Coltheart et al., 1977). This variable is of particular interest because it reflects the orthographic relatedness of a letter string with words in memory, and might therefore affect competition and inhibition processes in word retrieval. N can be computed for both words and pseudowords (in contrast to word frequency, for example), and indeed effects of N in behavioural tasks have been reported to differ between words and pseudowords. In the original lexical decision study of Coltheart et al. (1977), effects of N were reported only for nonwords, and were inhibitory (i.e. slower responses to more word-like nonwords which have a higher N). More recent studies have confirmed that rejections for pseudowords are slower for higher N's, but have also found facilitatory N effects for words (Andrews, 1989, Forster and Shen, 1996, Grainger and Jacobs, 1996, Sears et al., 1995). This pattern has also been found in the behavioural data of Holcomb et al. (2002), while their ERP data showed the same increase of N400 amplitudes with N for both words and pseudowords.

Opinions differ with respect to the origin of these orthographic neighbourhood effects. It has been suggested that N facilitates the lexical retrieval process at an early stage by means of feedback from the lexical to the letter level (Andrews, 1997), and that pseudowords with high N are more difficult to reject because they at least partly activate the corresponding word neighbours (Sears et al., 1999). In the multiple read-out model of Grainger and Jacobs (1996), it is assumed that lexical decisions are based on summed activation across all activated word representations. If a letter string activates many neighbours, the corresponding summed activation will be higher, which facilitates responses to words, but slows down rejections for pseudowords. In tasks where decisions are based on activation of individual representations, such as in word identification, the activation of neighbours inhibits the selection of the target item, and responses are therefore slower for higher Ns (e.g. Perea et al., 2004). This interpretation has been challenged by other researchers, who found N effects to depend on other task and stimulus properties. For example, N effects were not found for “No” responses in a semantic classification task (Forster and Shen, 1996), and have been reported to depend on the matching between word and non-word stimuli (Siakaluk et al., 2002). It is therefore still a matter of debate whether effects of N reflect fundamental lexical selection processes, or rather task-specific response strategies (see Andrews, 1997, Balota et al., 2004, Norris, 2006, for overviews).

Determining the time course and neuronal correlates of N effects in evoked brain responses may contribute to the understanding of processes modulated by orthographic neighbourhood variables. Early effects of this variable, i.e. in the latency range of the earliest previously reported effects of orthographic and lexical variables, would indicate that it indeed affects elementary word recognition processes. This argument would be strengthened if brain activation occurred in “classical” left-lateralised perisylvian language-related brain areas. Currently available neuroimaging data on N are as yet inconsistent. Binder et al. (2003), using a visual lexical decision task and orthographically matched words and pseudowords, did not find any brain areas for which activation significantly increased with neighbourhood size. Instead, they found that higher N produced lower activation to words in left prefrontal, angular and ventrolateral temporal cortex. In contrast, Fiebach et al. (2007) found differential effects of N for words and pseudowords in a lexical decision task in medial and mid-dorsolateral prefrontal cortex. Because these areas are commonly related to executive control functions rather than lexico-semantic processing, the authors argue that effects of N might arise only at a late post-lexical level. Consistent with this suggestion, the EEG study of Holcomb et al. (2002) revealed N effects around 400 ms after stimulus onset in a lexical decision task, which could reflect a later processing stage. In a semantic categorisation task, however, effects occurred earlier between 150–300 ms. We conclude at this point that effects of N are of great interest for psycholinguistic theories of lexical access, and that more data are needed to establish its effect on brain activation, and in particular its time course.

The investigation of N is complicated by the fact that N is negatively correlated with word length (e.g. Weekes, 1997). It is also positively correlated with typicality, i.e. more typically spelled words have more orthographic neighbours. The fMRI studies cited above controlled for word length and, where applicable, word frequency, but not orthographic typicality (Binder et al., 2003, Fiebach et al., 2007). Only the study of Holcomb et al. (2002) controlled all three of these variables. In our previous study (Hauk et al., 2006a), word length was negatively correlated with orthographic neighbourhood size and both variables were combined into a single predictor variable. Furthermore, results from the regression analysis were only presented for word stimuli, but not for pseudowords. In the current paper, we will therefore present a new analysis of the data reported by Hauk and colleagues, focusing on three novel questions: (1) Can we further characterise early brain responses by entering separate variables for N and word length into the regression analysis? (2) Can we reveal early differences between words and pseudowords for these variables, similar to those reported in the behavioural literature? (3) Do the source distributions differ between words and pseudowords, and how do they compare to existing neuroimaging results? In order to address these questions, we performed a new analysis on pseudoword data, and a re-analysis of our word data, including three variables: (i) word length measured as number of letters; (ii) orthographic neighbourhood size (N); (iii) orthographic typicality measured by bi- and trigram frequencies. Because converging results for the variable Typicality have already been presented in two independent studies (Hauk et al., 2006a, Hauk et al., 2006b), we will focus our analysis on the variables N and word length.