Introduction

Reading is a crucial cultural achievement, and efficient recognition of written words is at its core. Investigations of brain activations in response to visually presented words highlighted the role of a region in the left ventral occipito-temporal cortex (lvOT). This region is often also referred to as the visual word form area [1]. The presentation of letter strings (i.e., letter combinations that can be words or non-words) reliably activates this region [2–4], its structure and function is compromised in developmental reading disorders [5,6], lesions of this brain region result in severe reading deficits [7], and electrical stimulation of this brain region impairs word recognition [3,8]. However, there is no agreed-upon mechanistic understanding of the specific processes that lvOT contributes to visual word recognition [9]. Here, we propose a simple computational model of the lvOT function during reading. The lexical categorization model (LCM) integrates insights from neurocognitive and psycholinguistic research to explicitly model the profile of the lvOT responses to different types of orthographic stimuli.

The lvOT is part of the ventral visual processing stream [10]. It was proposed that lvOT receives converging bottom-up visual input from both hemispheres and that it processes abstract representations of recurring letter sequences–including sublexical units and small words [1,11]. This proposal is in part based on the finding that lvOT is sensitive to word-similarity [12,13], in the sense of decreasing lvOT activation (measured with functional magnetic resonance imaging; fMRI) with decreasing word-similarity of non-words. Specifically, high activation is observed when processing non-words containing letter sequences that frequently occur in real words (e.g., ‘ous’ in mousa). In contrast, non-words containing illegal letter combinations (e.g., mkzsq) result in low lvOT activation [12,13]. Seemingly contradictory, it was reported that more familiar (i.e., more frequently occurring) words elicit less lvOT activation as compared to rarely occurring (low frequent) words [14]. In sum, empirical data indicate that while word-similarity (in the sense of sub-lexical orthographic similarity) increases lvOT activation, word-familiarity (in the sense of word frequency) decreases lvOT activation. We interpret this counter-intuitive set of results as suggesting that lvOT responds in a non-linear fashion to the ‘word-likeness’ of orthographic strings, i.e., showing highest activity for words of intermediate word-likeness (e.g., words with low word-familiarity and non-words with high word-similarity) and least activity both for highly familiar, frequent words as well as for orthographically illegal and rarely co-occurring (‘word-un-like’) strings of letters (see also [4,15]).

The non-linear response profile of lvOT to different types of orthographic stimuli resembles the relationship between word-likeness and behavioral performance in word recognition tasks. Using a lexical decision task (categorical word/non-word decisions), Balota and Chumbley [16] observed that lexical decisions for letter strings with intermediate levels of word-likeness were more difficult (e.g., produced higher error rates) than decisions to very familiar words or very ‘word-un-like’ non-words (see also [17,18]). Based on these results, these authors proposed that letter strings at both extreme ends of the word-likeness continuum (i.e., highly familiar or highly word-un-like) can be classified quickly as words vs. non-words based on word-likeness information. In contrast, at intermediate levels of word-likeness, e.g., for rarely occurring words, words of a foreign language, or–as often used in psychological experiments–orthographically legal but meaningless pseudowords, uncertainty exists concerning the lexical nature of the letter strings. Balota and Chumbley [16] assume that this ambiguity is reduced by further analytic processing, e.g., based on processing sub-lexical units (e.g., letter) or word spellings.

We propose that lvOT implements, as core computation, an analogous process, i.e., a dichotomous lexical/non-lexical categorization of perceived orthographic strings informed by their word-likeness. lvOT thereby allows for fast access to subsequent linguistic processing stages for highly familiar words and, at the same time, filters out unknown or linguistically irrelevant percepts. Lexical categorization, thus, effectively functions as a gate that can save energy resources by preventing attempts to interpret unknown linguistic percepts. Thus, the proposed neurocognitive process of lexical categorization mirrors the behavioral pattern observed in lexical decision tasks [16], and the non-linear patterns of both lvOT activity and behavior are reflections of the level of uncertainty associated with the categorization process. The LCM proposal is consistent with one of the core computational functions of the ventral visual stream, i.e., categorizing percepts into different categories of objects based on a hierarchy of spatially different cortical processing stages [19]. For example, animate vs. inanimate objects activate separable ventral stream regions, while spatially dissociable sub-regions represent different lower-level features, such as faces or eyes for the animate subcategory. This object categorizing architecture is also known to categorize words vs. non-orthographic objects efficiently [8,20,21], and activation in these regions correlates with lexical decision performance [22]. Within the framework of hierarchically organized ventral stream processing, we thus assume that a categorical distinction between known and unknown orthographic percepts is computed after a lower-level categorization of sensory information into orthographic vs. non-orthographic input has taken place. More domain-general computational work has characterized the functional role of category-specific response patterns in vOT (e.g., [23]). In contrast, we here aim to more specifically model the effects of orthographic and lexical stimulus properties on word-selective cortex. Whereas previous work indicates that the former categorization process (e.g., words vs. faces) is based on visually distinct features, the lexical categorization process postulated here is proposed to rely on category-specific (e.g., orthographic, lexical) information.

To characterize the lvOT’s role for word recognition in a quantitatively explicit and testable manner, we implemented this hypothesis in a simple computational model: the lexical categorization model (LCM). The LCM predicts lexical categorization difficulty based on the word-likeness of the input letter string. Given its computational implementation, the LCM allows for model simulations and direct model evaluations by correlating simulated with empirical data. This proposal contrasts with previous conceptions of lvOT functioning in visual word recognition, which have been purely verbal-descriptive. Also, previous proposals divide into models that either assume linear or non-linear response profiles. Linear models include the local combination detector model [11,12] or the lexicon model [14,24,25]. The local combination detector model suggests that lvOT activation reflects a perceived letter string’s overlap with stored representations. The lexicon model assumes a word-frequency-based whole word lexicon search process (that lasts longer and thus elicits more activity in low-frequency words and non-words. However, these models cannot account for the non-linear response profile of lvOT discussed above. Non-linear models include the engagement and effort (E&E) model [4] and the interactive account (IA) model [15]. The E&E model assumes that the lvOT activation reflects the combination of the engagement and effort related to an orthographic stimulus. Engagement reflects that an illegal letter string (i.e., highly unfamiliar letter combinations) is less likely to be processed by a word recognition area when compared to a known word, so that brain activation is expected to be lower for the illegal string. Effort, in contrast, reflects that lvOT activation is greater for less often processed words than for frequently encountered words [4]. The IA model assumes orthographic processing based on the principles of predictive coding [26] based orthographic processing, such that lvOT activation reflects the error associated with internally generated predictions about the upcoming word [15]. The non-linearity of the IA model comes from the assumption that consonant strings do activate lvOT to a lesser extent since for these letter strings, few top-down predictions can be formed (e.g., in the context of an experimental task), which results in a low prediction error. In contrast, rare words are less often expected and thus often violate current predictions, so that the prediction error (and consequently the activation of lvOT), is high. Frequent words, on the other hand, can be predicted better so that both the prediction error and lvOT activation are lower for high- than for low-frequent words. None of these models of lvOT activation during word recognition have been implemented yet, so that explicit model comparisons are missing. Also, these alternative non-linear models can likely capture the response profile of lvOT activation to some extent but are conceptually harder to integrate into the predominant categorization-based accounts present for object recognition (e.g., [19] and see above) than the LCM.

Here, we evaluate the LCM against ad hoc implementations of the four alternative models introduced above and against a widely acknowledged, implemented cognitive model of visual word recognition, the Dual Route model [27] (which is in part a connectionist model). These evaluations focus primarily on comparing simulated vs. empirically observed lvOT responses to different stimulus characteristics–by generating simulated brain activations of observed activation patterns from different models (model-based fMRI analysis; see, e.g., [28]). First, models are compared based on their ability to simulate nine benchmark effects reported in the literature (Evaluation based on simulations). Second, we report empirical evaluations based on new fMRI studies that localize a neural correlate of the lexical categorization computation. After that, we compare models based on how well their simulations fit to the observed patterns of lvOT activatio (fMRI-based Evaluation). Finally, we investigate if training the core computation assumed in the LCM, i.e., the process of lexical categorization, increases reading efficiency (Training-based Evaluation).

Implementation of the lexical categorization model

To implement the LCM, we first derived the word-likeness distributions of a large set of orthographic strings: We estimated word-likeness (using an established measure from word recognition research, i.e., the orthographic Levenshtein distance, OLD20 [29]) of all German words with a length of five letters and an initial uppercase letter (i.e., nouns and noun-forms, e.g., plural form; N = 3,110; example: Augen/eyes), as well as for the same numbers of pseudowords (e.g., Augon) and consonant strings (e.g., Zbgtn) derived from these words. We visualized these distributions in Fig 1A. Expectedly, real words are characterized by high word-likeness, as shown by a relatively focused distribution between OLD20 values of 1 and 2 (gray distribution in Fig 1A). Technically, this indicates that they share many orthographic features with their twenty nearest orthographic neighbors. In contrast, the yellow distribution in Fig 1A, with its peak around an OLD20 of 3, demonstrates that consonant strings differ strongly from orthographically legal, existing words. For these two stimulus types with peaks at the extreme ends of the word-likeness distribution, word/non-word categorizations can be achieved with high degrees of certainty based on a word-likeness estimate. This certainty is reflected in the probabilities of a string being a word or a non-word given its word-likeness (gray and blue line, respectively, in Fig 1B). In contrast, the word-likeness (OLD20) distributions of words and pseudowords overlap strongly in the intermediate-likeness range (Fig 1A; consistent with previous literature [16]). Thus, for stimuli with intermediate word-likeness levels, lexical status is ambiguous, so that the word/non-word categorization cannot be based on word-likeness alone (probability of being a word or non-word is ~.5; Fig 1B). As a consequence, correct lexical categorization can only be achieved based on additional evidence [16], for example by comparing the presented letter string to more explicit representations that allow the correct spelling when writing words. To summarize, these considerations suggest that computing a lexical (word/non-word) categorization from an estimate of word-likeness is hard when word-likeness distributions of words and non-words overlap (as is the case for orthographically legal pseudowords) and easy when they do not (as is the case for frequent words as opposed to pure consonant strings). As Fig 1B (black line) indicates, this pattern of lexical categorization difficulty is very well captured by the information-theoretical concept of entropy, i.e., in our case, how much additional information is needed for the categorization (see Methods for details of the implementation). Thus, the basic assumption of the LCM is that the non-linear response profile of word-sensitive areas of lvOT should be reflected in the entropy of the lexical categorization, i.e., how difficult is it to categorize a letter string correctly.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Description of the lexical categorization model (LCM).

(a) Word-likeness distributions (kernel density estimates), based on the orthographic Levenshtein distance (OLD20 [29]) of words (gray), pseudowords (blue), and consonant strings (yellow) including an example for each category. (b) Probability that a letter string given its OLD20 value is a word (gray line) or a non-word (i.e., either pseudoword or consonant string; blue line). The black line represents the estimated entropy (see Eq 1, Methods section, for more details), which combines the probabilities of being a word or non-word across all possible OLD20 values. The LCM’s central hypothesis is that word-sensitive lvOT activation reflects this entropy function across all possible levels of word-likeness, effectively representing the difficulty of the lexical categorization process postulated for the lvOT. In an attempt to assess the internal stability of the LCM, we estimated LCM simulations based on only subsets of the words, pseudowords, and consonant strings used for the results presented here. In doing so, we found that only about 8% of the lexical items are needed to achieve stable LCM simulations (see S6 Fig).

https://doi.org/10.1371/journal.pcbi.1009995.g001

Results

Evaluation based on simulations: LCM simulations

As a first test, we used the LCM, as well as the implementations of four alternative neurocognitive models described above [1,4,15,30] and one purely cognitive model, the Dual-Route model [27], to assess how well these models reproduce lvOT activation reported in the literature for the most frequently published contrasts between different types of visually presented letter strings. We implemented the LCM simulations by transforming the word-likeness of each letter string into a lexical categorization uncertainty via the entropy function depicted in Fig 1B and described in more detail in the Methods section.

Fig 2A–2E displays LCM simulations for different types of letter strings (see Methods), all of which successfully reproduce published fMRI-based lvOT activation results. In detail, this involves fMRI contrasts of pseudowords > words [4] (Fig 2A); words > consonant strings [2] (Fig 2A); pseudowords > words > consonant strings [31] (Fig 2A); pseudohomophones > words and pseudohomophones = pseudowords [24]; pseudowords > words matched on multiple lexical characteristics [32] (Fig 2B); the word similarity effect: low word similarity < intermediate word similarity < high word similarity = words [12] (Fig 2C); increasing lvOT activation with decreasing word frequency [14] including pseudowords (Fig 2D; note that when only words were used, the beta-weight in that study was reduced from -0.40 to -0.17; see also [33–35]); bigram frequency effect: increasing lvOT activation with increasing bigram frequency [13] (Fig 2E; both frequency measures had been log. transformed). Note that bigram frequency was also previously used for estimating word-likeness and is correlated to the OLD20 parameter (e.g., see Fig 2B in Ref. [18]). Therefore, we also visualize the non-linear bigram frequency effect in the LCM simulation (Fig 2E, dashed line), resolving the inconsistent findings concerning lvOT and bigram frequency previously reported, i.e., an increase in activation with bigram frequency in [13] as opposed to a decrease in [35].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Evaluation based on simulations: LCM and model comparisons based on simulations of lvOT benchmark effects from the fMRI literature.

Model comparisons involve the lexicon model, local combination detector model (CD), effort and engagement model (E&E), the interactive account model (IA; for implementations and detailed simulations see S1 Text), and the cognitive Dual Route model, specifically its orthographic lexicon (DRC Lex.) and its grapheme to phoneme conversion route (DRC GPR; for detailed simulations see S1 Fig). Simulated lvOT activation (in arbitrary units: min = 0; max = 6) for all groups of letter strings is presented using bar graphs depicting their respective mean activation. In addition, horizontal black bars indicate significant differences of the simulation results between letter string categories, as derived from linear models (Bonferroni corrected). LCM simulated lvOT activation is presented, from left to right, (a) for words (W), pseudowords (PW), consonant strings (CS), pseudohomophones (PH), (b) words and pseudowords matched on number of syllables, number of Coltheart’s orthographic neighbors, frequency of the highest frequency neighbor, initial bigram frequency, final bigram frequency, and summated bigram frequency (mW, mPW), and (c) the word similarity effect comparing words (cmW: comparative matched words) to non-words with high word similarity (matched on quadrigram frequency; hWS), to non-words with intermediate word similarity (matched on bigram frequency; iWS), and, to non-words with low word similarity (lWS). In addition, (d) the effect of log. transformed word frequency (for all words and pseudowords as tested in [14]) and (e) log. transformed bigram frequency are presented as scatter plots with a linear regression line. Note that for bigram frequency, also a non-linear regression line is shown. In (d) and (e), each dot represents one letter string, the more saturated the blue gets, the more letter strings are included. See text for more detailed description of the replicated benchmark effects including the specific stimulus sets used and references to the original studies. (f) Qualitative model comparisons showing the sum of correctly simulated stimulus differences (orange bars) and all correct minus all incorrectly simulated effects excluding null effects (cyan bars). The LCM and the IA were able to correctly simulate all contrasts. (g) Correlation matrix of all model parameters included in the model comparison, showing that simulations of the non-linear models, i.e., LCM, E&E, and IA, were substantially correlated (all r’s > .4).

https://doi.org/10.1371/journal.pcbi.1009995.g002

When we simulate these contrasts with all five neurocognitive models and the German version of the Dual Route model, the different models can be compared based on the number of correctly simulated contrasts (Fig 2F and 2G; for details on the implementations and the results of the simulations of alternative models [1,4,14,15,27], see S1 Text). Note that for the Dual Route model [27], we conducted two such simulations, one reflecting the model-based activations of the orthographic lexicon, similar as assumed in the lexicon model [30], and one reflecting the activation of the grapheme-to-phoneme route, as suggested by [25] (reflecting letter-by-letter decoding). This differentiation allowed us to investigate the model’s two routes as independent hypotheses for the cognitive processes implemented in the word-sensitive lvOT (see S1 Fig for details). We also investigated the response times simulated by the Dual Route model. However, no difference between the response times and the Dual Route model’s orthographic lexicon simulation was found, so that this was not included separately into the model comparison.

Only the LCM (Fig 2A–2E) and the IA model (S1 Text) could simulate all nine contrasts of interest correctly (Fig 2F). The E&E model, also implementing a non-linear response profile, performs better (6 out of 9 contrasts) than the neurocognitive models assuming a linear response profile and the cognitive Dual Route model. This finding suggests that implementing a non-linear response profile is a critical feature of the lvOT response to letter strings. This conclusion receives further support when we penalize the model performance for incorrect simulations (minus one for each incorrectly simulated effect) and from the positive inter-correlations among LCM, IA, and E&E models (Fig 2G; all r’s > .4). Given these results, we focused our empirical fMRI-based model comparisons on the three best-performing models, i.e., LCM, IA, and E&E models.

fMRI-based evaluation: Brain activity measured with functional MRI

To test whether the quantitative predictions of the LCM capture empirically measured BOLD activity in lvOT, we used data from three fMRI studies (cf. Methods for acquisition parameters and preprocessing). In this section, we will first focus on the LCM effect and model comparisons. After that, we will present two additional effects of interest, i.e., the correlation of brain activation and word-likeness (OLD20) and the regions with higher activation levels for words, compared to pseudowords.

fMRI Study 1.

Participants (N = 15 viewed a set of letter strings covering a wide range of word-likeness (i.e., words, pseudowords, and consonant strings; each five letters length), and strings of scrambled letters as a low-level perceptual control condition in a blocked design with 6 blocks of 16 items for each of the four conditions, presented in a random sequence (see Methods for details). Participants pressed a button whenever they detected a target stimulus (‘#####’). Control stimuli with scrambled letters were generated by randomly exchanging 90% of the pixels of the monochrome images of the word condition that were used for stimulus presentation, to obtain a list of maximally unfamiliar stimuli with identical physical input as the word condition. The analysis of BOLD activation was exclusively based on the lexical categorization difficulty as simulated by the LCM. In detail, we used item-specific LCM simulations as a continuous predictor for the fMRI data without explicitly accounting for condition differences and treated the control condition as having a very low word-likeness so that the LCM simulation resulted in a value of 0 (see Methods for details). We reasoned that if the proposed categorization process is indeed implemented in lvOT, quantitative LCM simulations should predict activation in this area. Our results strongly support this prediction: Across the entire brain volume, only voxels in the lvOT show a positive relationship between LCM simulations and BOLD response (Fig 3A; see also Fig 4, left-most column, and Table 1). Thus, the higher the categorization uncertainty expressed in the entropy derived from the word-likeness distributions from the lexicon, the higher the BOLD response in the identified lvOT voxels. The location of the detected activation cluster is consistent with previous functional localizations reported for the visual word form area (including published MNI peak coordinates such as -48, -56, -16 [12]).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. fMRI-based Evaluation of the LCM model: fMRI whole brain analyses.

(a, b, d, e) Significant correlation results between BOLD signals and LCM simulations modeled as a single, continuous predictor in Study 1 (a) and Study 3 (d), and between BOLD activation and word-likeness represented by the OLD20 as a single, continuous predictor in Study 2 (b) and Study 3 (e). (c, f) Significantly increased BOLD signals for words as compared to non-words in Study 2 (c) and Study 3 (f). For OLD20 results and word > non-word contrasts of Study 1, see S2 Fig). Thresholds for all whole brain analyses: voxel level: p < .001 uncorrected (cluster-forming); cluster level: p < .05 family-wise error corrected. No other regions than those displayed were significant.

https://doi.org/10.1371/journal.pcbi.1009995.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Comparison of BOLD activation across studies and contrasts.

Significant LCM activation clusters in the lvOT across studies show some overlap, but no overlap to word-likeness and lexicality contrasts. Thresholds for all whole brain analyses: voxel level: p < .001 uncorrected (cluster-forming); cluster level: p < .05 family-wise error corrected.

https://doi.org/10.1371/journal.pcbi.1009995.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Significant activation clusters from the fMRI evaluation with respective anatomical labels (most likely regions from the Juelich and Harvard-Oxford atlases including % overlap), cluster size (in mm³), and peak voxel coordinates (MNI space).

https://doi.org/10.1371/journal.pcbi.1009995.t001

The model simulations for the two other non-linear models (IA, E&E) conducted with the identical set of stimulus materials resulted in qualitatively similar patterns (Consonant strings/CS < words/W < pseudowords/PW) as for the LCM model (see simulation results in Fig 5A–5C). However, critical quantitative differences remained, including much larger differences between modeled activations for words and pseudowords than predicted based on the LCM (compare Fig 5B and 5C to 5A). The empirically measured BOLD activation at the peak voxel of the lvOT cluster (Fig 5J) also showed the same qualitative pattern of differences as simulated by the three models, i.e., CS < W < PW (cp. Fig 5A–5C). To implement a quantitative comparison between the three models, we transformed all simulations and the empirical BOLD data to a common scale (i.e., z-transformation within each dataset to conserve relative differences but remove absolute differences). Then we compared every model’s contrast differences with the empirically measured contrast differences (i.e., W vs. PW, W vs. CS, PW vs. CS). As a difference measure, we set each contrast difference in relation to the standard deviation from the observed data (see Fig 5M; Fig 5P for all individual contrasts), allowing a quantitative comparison of the predicted vs. the observed data. As Fig 5M shows, quantitative predictions of the LCM accounted best for the empirical lvOT ROI data, yielding the lowest difference between predicted and observed data summed across all three contrasts (i.e., <1 SD). When examining these differences by contrasts, LCM predictions were best for the W vs. PW and PW vs. CS contrasts. In contrast, the E&E model was more accurate for the W vs. CS contrast (but the LCM was still within 2 SDs of the empirically measured activation difference between W and CS). In summary, the fMRI data of Study 1 indicate that the LCM provides a fair characterization of BOLD activation patterns in several lvOT voxels during visual word recognition and that it predicts activity in these voxels better than other current models of lvOT function.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5.

Simulations for Words (W), pseudowords (PW), and consonant strings (CS; all medians) used in Study 1 from (a) the lexical categorization model (LCM), (b) the interactive account model (IA), and (c) the effort and engagement model (E&E). (d) LCM, (e) IA, and (f) E&E simulations for Study 2. (g) LCM, (h) IA, and (i) E&E simulations for Study 3. Empirically observed lvOT ROI activation for W, PW, CS, and scrambled letters (SL), extracted from the same peak voxel region, i.e., defined in Study 1, of interest (ROI) of (j) Study 1, and (k) Study 2. (l) For the animal decision and catch trial detection task of Study 3 the ROI data was extracted from the peak voxel of Study 3. We present the percent signal change, in arbitrary units, for each condition, including the variance across participants (horizontal line represents the median; box +- 1 standard deviation; whiskers +- 2 standard deviations). Besides, we present model comparisons for the LCM, E&E, and IA models of Study 1 (S1) in (m), of Study 2 (S2) in (n), and of Study 3 (S3) in (o). We show the mean difference between the model simulated contrasts and the observed contrast differences from the ROI data. We standardize these model comparisons via the standard deviations of the observed data (SD). I.e., the difference between simulated and observed contrast differences in standard deviations of the observed data. For Study 1, we summarize three contrasts, which can be inspected in detail in (p), Study 2 includes only one contrast presented in (n), and Study 3 combines six contrasts shown in detail in (q). In the single contrast figures (pq), the solid black line and dot show the mean observed difference (across all participants), dashed lines +- 1 standard deviation, and dotted lines +-2 standard deviations also from the observed differences. Green dots show the LCM simulated contrast estimate, blue dots the IA contrast estimate, and red dots the E&E contrast estimates. In (q) the left panel represents the animal detection and the right panel the catch trial detection task. (r) The linear relationship between lvOT ROI activation and LCM-simulations in study 3 separated for words and non-words and the two tasks conducted in the study.

https://doi.org/10.1371/journal.pcbi.1009995.g005

Study 2.

In Study 1, letter strings were presented in condition-specific blocks, so that the predictability of the next item concerning word/non-word categorization was high. Since the LCM does not account for the experimental context, it cannot be excluded that the blocked design may have strategically reduced the amount of processing devoted to the lexical categorization. To exclude this potential confound, we examined a second fMRI study in which stimulus categories (words and pseudowords matched for word-likeness/OLD20) were randomly intermixed (event-related design). Participants (N = 39) performed a silent reading task and had to detect a rarely occurring target word, i.e., the German word Taste (button). The correlation analysis from study 1, i.e., including the LCM simulated activation, did not result in any significant activation cluster. This finding was not surprising since stimuli had been selected to equalize the word-likeness between words and pseudowords with low variance.

Still, with this stimulus material, model simulations differed between the three models so that we were able to implement the model comparison analysis using the data from the peak voxel of study 1. The LCM simulations predicted high lvOT activation for both stimulus conditions with only subtle condition differences (Words>Pseudowords; Fig 5D). In contrast, both the IA and E&E models predicted substantially higher activation for pseudowords than words (Fig 5E and 5F). In the peak voxel from study 1, we found positive activation levels with virtually no differences between words and pseudowords (see Fig 5K). Model comparisons based on the word/pseudoword contrast showed that the LCM predicted the observed activation contrast best (i.e., <1 SD; Fig 5N). To summarize, here we replicated the superior prediction performance of the LCM model in an event-related design, even when no stimulus-specific difference was found in a ROI located in the lvOT.

Study 3.

In study 1, we used stimuli with characteristics similar to previous studies. In study 2, in contrast, we included words and pseudowords with equalized word-likeness to test the LCM model in a situation in which it predicted a different pattern than alternative models. In study 3 (N = 35), we aimed at replicating this result in a study that uses stimulus materials resulting in an LCM activation pattern previously unseen in the literature and that could not be captured by the alternative models. To achieve this, we again investigated words, pseudowords, and consonant strings. We selected stimuli so that the LCM-simulated lvOT activation predicted a pattern that was inverted relative to the results predominant in the literature and observed in our study 1, i.e., with greatest activity elicited by consonant strings (CS > W > PW; cf. Fig 5G and compare to the more typically observed pattern of PW > W > CS shown in Fig 2A and 2B). If word-sensitive regions in lvOT indeed implement a word similarity-based lexical categorization process, lvOT voxels should be detectable that show this pattern of activation. Contrary to this LCM prediction, the simulations of the IA and E&E models (Fig 5H and 5I) predict different activation patterns for this specific stimulus set (IA: PW > W > CS; E&E: PW > CS > W), of which only the IA model reflects the ‘classic’ pattern predominantly found in the literature (i.e., when using ‘conventional’ stimulus sets). Second, to assess whether the depth of processing of the participants’ task (i.e., word recognition vs. semantic analysis) influences the assumed lexical categorization process, two different tasks were included in this study, an animal detection task (AD) and a catch trial detection task (CD). Each participant conducted two runs of each task, with two task sequences (AD/CD/AD/CD vs. CD/AD/CD/AD) assigned randomly to participants.

Analogous to Study 1 and 2, we used the LCM simulated activation as a single continuous predictor in a first whole-brain analysis. Again, a robust correlation of BOLD activation and LCM simulations was observed in a lvOT cluster (Fig 3D), slightly more laterally and posteriorly than in Study 1 (cp. columns 1 and 2 in Fig 4) but with a peak within 1 cm distance of the lvOT peak in Study 1 (Table 1; LCM effect Study 3, Cluster 3). We also found correlations between LCM simulations and BOLD activation in the left and right parietal cortex (Fig 3D and Table 1). Peak voxel activation of the lvOT cluster showed significant LCM correlations in both words and non-words (PW and CS combined) and both tasks (all t’s(34) > 4.3; all p’s < .001; Fig 5R). For this analysis, we estimated the activation for words and non-words separated for three LCM activation levels for each task (see blue and gray dots in Fig 5R). In this region of interest analysis, we did not find an interaction of condition and task (all t’s < 1.47; all p’s > 0.14) and no significant differences between stimulus categories (all t’s < 1.34; all p’s > 0.17; Fig 5L). Still, we found a task effect showing a higher activation in the animal detection task (t(34) = -2.2; p < 0.03). This finding is consistent with the behavioral results, which showed that the animal detection task was more demanding than the catch trial detection task, reflected in more errors (AD: 4.4%; CD: 1.5%; GLM Estimate: -0.03; SE = 0.015; t = 2.2) and longer response times (AD: 644 ms; CD: 520 ms; LM Estimate: -0.2; SE = 0.02; t = 9.7).

The model comparisons, implemented as in Studies 1 and 2, found that the LCM had, again, the lowest error in predicting the empirically observed pattern of activation contrasts (Fig 5O). Like in Studies 1 and 2, the E&E model simulated the lvOT pattern better than the IA model. In detail, of the six contrasts investigated, the LCM had three predicted contrasts within one standard deviation (SD) of the empirically observed BOLD activation difference, two that fell within 2 SDs, and one that was just above 2 SDs (Fig 5Q). In contrast, the E&E (two < 1 SD’s & four > 2 SD’s) and the IA (two < 2 SD’s & four > 2 SD’s) simulations resulted in larger differences. To protect against potential biases due to the ROI definition from the LCM contrast, we explored whether the alternative E&E and IA models predict brain activation in other lvOT voxels. To this end, we implemented two separate additional whole-brain analyses in Study 3 that correlate E&E and IA model parameters with BOLD activation. As the results in S2 Fig show, we observed no significant voxels in lvOT, neither for the E&E nor for the IA model. However, BOLD activations in frontal and bilateral lateral temporal regions correlated with the IA model simulation and left peri-central activation was associated with the E&E model parameter. Thus, the findings of Study 3 further support the proposal that involvement of lvOT subregions during visual word recognition is best described by a lexical categorization computation as implemented in the LCM.

Studies 1–3 –Word-likeness and Lexicality effects.

Besides the LCM-based analyses, we found correlations with word-likeness (i.e., represented by OLD20; see above) in all three studies. We localized the word-likeness representations predominately to areas posterior to LCM effect in lvOT, i.e., to the posterior lvOT effect in Studies 1 (S3 Fig) and 2 (Figs 3B and 4 column 3) and parietal cortex (Fig 3E, see also Table 1). In Study 3, we also observed a correlation between OLD20 and BOLD signals in the sensorimotor cortex. A dichotomous lexicality effect with larger activity for words (W) than non-words (PW and CS), was found most consistently in brain regions downstream to lvOT in the inferior frontal cortex (Fig 3C and 3F and Table 1). In Study 2, we found additional activation for words in the Superior frontal gyrus (Fig 3C). In Study 3, we observed activation increases for words in the left superior temporal gyrus and extensive activity in the left and right occipital cortex. (Fig 3F).

In sum, empirical data from three fMRI studies support the proposal that there is a region in lvOT that implements a lexical categorization process which operates upon an estimate of word-likeness computed by upstream, more posteriorly located brain regions. Within the here-proposed LCM model, this lexical categorization process is modeled quantitatively as an entropy measure reflecting the difficulty of lexical categorization given the input’s word-likeness (see Fig 1), and LCM simulations of three different stimulus sets have accounted better for lvOT activation patterns than two alternative models. Of note, this LCM-based prediction of lvOT activity was obtained without restricting the search space to this area, but rather in whole-brain analyses, whereas only in Study 2, where no activation difference was expected due to stimulus characteristics, the LCM analysis was based on ROI data. To avoid bias, we demonstrate with further whole-brain analyses in Study 3 that competitor models do not account for lvOT activity (S2 Fig). Also, word-likeness accounted for BOLD activation differences in brain regions posterior to the LCM effect, which is consistent with the proposal that word recognition processes in lvOT operate upon a previously computed word-likeness estimate. Furthermore, we found dichotomous lexicality (Words > Non-words) effects anterior to the LCM effect (e.g., in left inferior frontal regions), which is consistent with a further assumption of LCM, i.e., that the lexical categorization process ‘selects’ orthographic input for further linguistic processing. In summary, we conclude that the empirical evaluation based on three fMRI datasets provides support for the LCM.

Training-based evaluation

After repeatedly finding that lexical categorization difficulty is useful to describe the activation pattern of a brain area identified as highly relevant for reading, we tested the causal role of the assumed cognitive process for reading. We reasoned that when the postulated cognitive process of lexical categorization is relevant for reading, modifications of this process should affect reading performance, i.e., here defined as reading speed (which is often used as performance measure in German, see [36]). To test this hypothesis, we implemented a training procedure for 76 non-native early-stage learners of the German language who were from diverse language backgrounds. The training started and ended with a reading speed assessment (Fig 6A). Inbetween, three training sessions of about 50 min in length were implemented on consecutive days. To train the lexical categorization process, we used a lexical decision task, i.e., the task we motivated our model from [16]. Previous research also found that performance in a lexical decision task could predict lvOT activation [22]. To reinforce lexical categorization behavior, we implemented feedback that rewarded correct lexical decisions with a green square and indicated incorrect responses with a red square (Fig 6B). For comparison, we included a lexical decision dataset with the same stimuli (800 words and non-words) and a task from a sample of native German readers (N = 48) from one session.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Training-based Evaluation.

Behavioral training of lexical categorization: Paradigm and results. (a) Temporal structure of the training study, including the three training sessions on three separate days, as well as pre- and post-training assessments of reading speed (based on the Salzburger Lese Screening/SLS; see Methods). (b) Lexical decision paradigm performed by the participants during the three training sessions, including feedback screens. (c, d) Lexical decision response times in relation to (c) word-likeness (OLD20; i.e., high OLD20 represents low word-likeness; see Fig 1) and (d) separated for stimulus categories (W: Words; PW: Pseudowords; CS: Consonant strings) from 76 non-native learners of German across three sessions. For comparison, data for a group of 48 native German speakers (i.e., typical readers) who completed the lexical decision task with the same stimuli are presented. (e) Predicted response time reduction with session, i.e., from session 1 to 2 (green) and from session 1 to 3 (blue), extracted from fitted linear mixed models using the remef function to remove all but the variance which reflects the session by entropy interaction (for details see the Methods section). (f) Change of reading speed (pre-post; see Methods) correlated with the individual training effect (quantified as the estimate for the LCM by session interaction from a linear mixed model with random slopes). The boxplot (right panel) shows the overall increase in reading speed on the pre-post test. For all boxplots, the horizontal lines represent the median, the boxes represent +/- 1 standard deviation, and the whiskers represent +/- 2 standard deviations.

https://doi.org/10.1371/journal.pcbi.1009995.g006

The analysis of response times from lexical decisions showed a non-linear effect of word likeness (i.e., OLD20) for native and non-native readers, with longest response times in an intermediate range of word-likeness as predicted by the LCM (Fig 6C; see also [16]). For the non-native readers, we found a significant interaction of the lexical categorization difficulty predictor derived from the LCM and training session, indicating that lexical categorization becomes more efficient with training. This interaction effect is most obvious for intermediate levels of word-likeness levels, which represent mostly words and pseudowords and which is characterized by categorization difficulty due to overlapping word similarity distributions (Estimate: -0.03; SE = 0.01; t = 2.4; See Fig 6D and 6E). Note that to model this non-linear relationship between OLD20 and response times, we used the first and second polynomials, i.e., the linear and squared OLD20.

Response times also became faster from session to session (main effect; Estimate: -0.04; SE = 0.005; t = 8.0). However, non-native readers only reached the performance of native German readers when processing non-words (see Fig 6D; equivalence test for a subset containing only pseudowords from session three and the data from typical readers: Upper bound: t(122) = -2.3; p = 0.012; Lower bound: t(122) = 3.2; p = 0.001; for consonant strings: Upper bound: t(122) = -2.6; p = 0.005; Lower bound: t(122) = 2.8; p = 0.003). For words, no significant group differences could be found (t(122) = 1.7; p = 0.085) but also equivalence could not be established (Upper bound: t(122) = -1.0; p = 0.166; Lower bound: t(122) = 4.5; p < 0.001), as we did not find a significant difference from both equivalence bounds. Still, note that the comparisons of word-related response times from sessions 1 and 2 between language learners and typical readers resulted in significant differences (both t’s > 3.7; p’s < .001).

Next, we implemented a linear mixed effect model that included estimating the random slope of the interaction term (i.e., the lexical categorization difficulty by training session interaction). Including this parameter in the model describing the reaction times in the training allowed us to estimate the increase of the lexical categorization effect with training for each non-native reader. This individualized estimate represents how strong the quadratic lexical categorization effect increased with training. This predictor allowed us to determine if the LCM-based training effect transfers to an increase in reading speed (determined using the pre/post-training reading speed test; see Methods). We estimated a regression analysis with reading speed change as the dependent variable and the effect estimates of the individual lexical categorization by session interaction effects, the pre-training reading speed, and their interaction as independent variables. We found no significant interaction (GLM Estimate: -25.8; SE = 13.4; t = -1.9; p = 0.058) but significant main effects of pre-training speed (GLM Estimate: -1.1; SE = 0.5; t = -2.3; p = 0.022) and of the individual lexical categorization by session interaction effects (GLM Estimate: 416.7; SE = 205.4; t = 2.0; p = 0.046; see Fig 6F) on the pre-post change in reading speed. Overall, on average, the increase in reading speed after the lexical categorization training was 20% (t(75) = 4.5; p = 0.001; Fig 6F right panel). In sum, this final evaluation of the LCM suggests that the process of lexical categorization contributes to efficient reading. This finding supports the LCM’s assumption of lexical categorization as a core process of visual word recognition.

Discussion

In the work reported here, we used model-based fMRI analysis to clarify the nature of computations implemented in (subregions of) left-ventral occipitotemporal cortex (lvOT) in response to orthographic percepts (letter strings). To achieve this, we implemented the four most prevalent theories of lvOT function during language processing (e.g., [11]), the most widely accepted psychological model of visual word recognition (i.e., the Dual Route Model; [27]), and a new hypothesis, the lexical categorization model (LCM), as computationally explicit models and performed direct model comparisons using different types of data. The newly proposed LCM model postulates that word-sensitive lvOT areas in the vicinity of the functionally localized visual word form area implement a categorization process that classifies letter strings into meaningful (words) vs. meaningless input, preventing attempts to apply downstream lexical-semantic processing to non-linguistic input. The LCM implementation makes explicit how the lexical categorization process is related to neural processing, i.e., via the estimated difficulty of categorizing a letter string.

We compared the implementations of the LCM and alternative models directly by simulating established effects from previous fMRI studies. Through its lexical categorization mechanism, the LCM successfully accounts for the intricate pattern of lvOT activation in response to a variety of different (sets of) orthographic stimuli, irrespective of their specific task context (compare, e.g., [12,14]). Most notably, the LCM model can simulate multiple published fMRI findings that have been viewed as contradictory, and it does so better than competing models. In addition, we found that alternative models implementing processes that in principle allow for non-linear activation patterns (IA model, [15]; E&E model, [4]) performed better than linear models like the local combination detector model (e.g., [11]). Furthermore, we demonstrate that the LCM reliably predicts empirically measured BOLD fMRI activations in voxels of the mid-lvOT and outperforms alternative non-linear models in three experiments that differed concerning the selection and composition of stimulus sets (in both whole-brain and peak-voxel defined ROI analyses). We observed that word-likeness, i.e., the quantitative psycholinguistic property on which the LCM’s lexical categorization process is based, is predominantly represented in regions posterior to the brain regions showing a LCM effect. In contrast, a dichotomous lexicality effect with increased activity for words is implemented in brain regions anterior to the lvOT subregion associated with the LCM effect. These findings are consistent with the proposal that lvOT’s categorization process gates known words into subsequent stages of lexical processing (see also [37]).

Finally, we demonstrate that repetitive behavioral training based on lexical decisions, which involve lexical categorization as a core process, significantly increases reading speed. Given that reading speed as tested in this study does not explicitly involve word/nonword decisions, we argue that this training study can provide causal evidence for an involvement of lexical categorization in the process of reading. Overall, these findings suggest that during reading, following the initial stages of visual-perceptual processing, posterior brain regions compute an estimate of the word-likeness (relative to the known lexicon of words). According to the LCM model, this word-likeness estimate is fed forward to a lexical categorization process implemented in a subregion of lvOT, which in turn precedes the extraction of word meaning (lexical access) at downstream cortical sites, including more anterior temporal and frontal cortex of the left hemisphere (Fig 7).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7.

Schematic description of processes during visual word recognition as assumed in the lexical categorization model. The figure includes (i) word-likeness estimations in posterior visual-perceptual regions, (ii) lexical categorization in the left ventral occipito-temporal cortex (lvOT), and (iii) the extraction of word meaning in anterior regions including temporal and inferior frontal cortex. The lexical categorization process implemented in lvOT is schematically represented by the uncertainty (entropy), to visualize the LCM’s assumption that higher degrees of categorization uncertainty–in sections of intermediate word-likeness (yellow)–may require further elaborative processing to reach a lexical categorization.

https://doi.org/10.1371/journal.pcbi.1009995.g007

The LCM is a quantitatively explicit model of the role of lvOT in visual word recognition. It allows for explicit comparisons to alternative models, given a specific set of stimuli. Notably, the LCM outperforms, both concerning simulating published data and modeling empirical data, multiple alternative models [4,11,15,27,30], most of which have been implemented computationally here for the first time. The comparisons based on model simulations of published BOLD activation effects showed a clear advantage for models that assume a non-linear response profile of the lvOT, i.e., LCM, the interactive account [15], and the engagement & effort model [4]. Among these three models, the LCM performed best in our fMRI data-based model comparisons.

Our work differs from previous approaches by considering a wider range of orthographic word-likeness (i.e., ranging from frequent words to pseudowords to consonant strings), whereas previous studies mostly focused only on parts of the word-likeness distribution (e.g., high vs. low-frequency [14,33,34] or words vs. pseudowords [31,32]). Also, we used a more optimal word-likeness estimate (i.e., the orthographic Levenshtein distance/OLD20 as compared to, e.g., quadrigram frequency), as is evident from previous comparisons of different word-likeness measures based on behavioral data [29]. We complemented this investigation with LCM simulation studies using alternative word-likeness estimates (see S4 Fig). The OLD20 word-likeness estimate is theoretically meaningful, as it is a reasonable proxy for the perceptual familiarity that one acquires while becoming an efficient reader [38]. On the other hand, we have argued in previous work that OLD20 may lack neuronal plausibility as it potentially subsumes multiple cognitive processes (pre-lexical and lexical; e.g., see [18]), as suggested by our finding that OLD20 correlates not only correlates with other word-likeness characteristics but also with lexical frequency [18]. To explore this in somewhat more depth, we have complemented the present investigation with LCM simulations based on three widely accepted alternative estimates of orthographic word-likeness, i.e., Coltheart’s N, trigram frequency, and quadrigram frequency. As S4 Fig shows, OLD20 captures the lvOT activation patterns best (in terms of correctly predicted effects). Also, one could criticize that OLD20 estimations are based exclusively on a lexicon of words while in the context of the LCM, they are also used to model brain responses to non-words. However, we think that this plausibly mirrors processes during visual word recognition, which have been shaped based on the words we have learned. The processes assumed by the LCM should, thus, also function at different stages of development (i.e., based on different lexicon sizes albeit at varying levels of precision). The effects of different lexicon sizes on LCM simulations are reported in S5 Fig. As can be seen, words, pseudowords, consonant strings, and pseudo-homophones result in the expected pattern (PW = PH > W > CS) only after at least 10,000 words (~ 5% of the full lexicon) were included in the lexicon (i.e., which was the basis for the OLD20 estimation). In this context, it is also noteworthy that stable LCM simulations can be achieved with lexicon sizes as small as ~ 8% of the lexicon used in the present work (S6 Fig). The LCM’s entropy estimation not only reflects how our brain responds to words, but also how non-words are processed. We think that including non-word distributions is essential, as the empirical tasks that the model simulates included both words and non-words. S1 Text demonstrates that an LCM without OLD20 distributions for non-words cannot reproduce core benchmark findings from the literature (like greater activity for pseudowords as compared to words). However, expectations of the reading material presented in different task contexts may shape the cognitive processes engaged. We propose that these and similar issues involving task, stimulus, and lexicon effects on lvOT activation can be systematically investigated in future work using the quantitative approach introduced here.

At its core, the LCM-simulated activation reflects the difficulty of the process of lexical categorization based on the word-likeness of the orthographic input. It is easy to identify a letter string as a meaningful word when the word-likeness (i.e., orthographic familiarity) is high. It is also easy to categorize a letter string as meaningless when early visual processing stages identify it as unfamiliar given what we have learned, i.e., the lexicon of words we have acquired so far. Both describe efficient cases of lexical categorization of letter strings, resulting in fast responses. At intermediate levels of word-likeness, however, lexical categorization difficulty is high. The LCM postulates that this results from similar orthographic-perceptual familiarity of words and non-words fed into (word-sensitive subregions of) lvOT from lower levels, making a more detailed analysis of the orthographic input necessary. In the LCM implementation, this is accounted for by the partly overlapping distributions of word-likeness (OLD20) of words and non-words. The additional processing required by letter strings with intermediate word-likeness results in increased lvOT activation and slower responses, a proposal consistent with seminal behavioral work which proposed spelling information as a potential source aiding categorization when word-likeness information is insufficient [16]. This, in turn, is compatible with our observation from fMRI Study 3, in which we used stimulus materials that were particularly difficult to categorize (e.g., Consonant strings with a high word-likeness). In this experiment, we observed that a parietal network that likely represents a dorsal sensorimotor interface connecting auditory-phonological with articulatory representations [39] co-activates with lvOT when lexical integration difficulty increases. This parietal co-activation, interestingly, also relates our results to the recent proposal of a posterior/anterior dissociation within the word-selective ventral occipito-temporal cortex [37,40]: Whereas posterior lvOT (also referred to as VWFA-1 by [37,40]) has been characterized as implementing pre-lexical processing during visual word recognition and was shown to be structurally connected to the parietal cortex [37,41], the anterior lvOT (VWFA-2) shows selectivity to lexical stimulus properties [37,40]. The LCM-correlate identified in lvOT in the present study (Study 1: y = -64; Study 3: y = -75) is located close to the posterior lvOT subregion (Ref. [37]: y < -63; Ref. [40]: y ~ -69), suggesting a primarily prelexical nature of the categorization process implemented in the LCM. A proposal that is in line with the functional profiles of reading-related brain regions identified with EcOG electrodes [3,42]. Note, however, that these group-level effects also need to be interpreted in light of substantial inter-individual differences in functional localization of word-selective activity (e.g., [41]), so that the association between LCM effects and posterior lvOT/VWFA-2 should be replicated based on individual functional localizers.

Various computational models have described cognitive processes in higher vision implemented in the vOT. These models are concerned with visual word recognition (e.g., [23,43]), object or face recognition (e.g., [44–46]). Current computational cognitive model approaches are often based on deep neuronal network models (DNNs) that can investigate complex system behavior and differences in computational architectures [47,48]. Also, these models achieve highly accurate predictions for vOT activation patterns during object recognition (e.g., [49]). Still, DNNs typically include numerous free parameters fitted in a specific training setup, i.e., typically in an image classification task. This characteristic of DNNs comes at the cost of reduced model transparency and explainability (e.g., [47,50]). More transparent approaches, like the Template model used by Yeatman and Kay [23], have a similar scope, i.e., an image input and a categorical output (word vs. face), but achieve this with only a limited number of free parameters (N = 3). The model described here shares this aspect of greater transparency and explainability. Developing the LCM was under the premise that the model should be as simple as possible in the implementation and as deterministic as possible (i.e., with a minimal number of free parameters). Given that the LCM model output only depends on the words stored in the lexicon, it is deterministic in the sense of not including any free parameters and thus also fully transparent. However, the current implementation of LCM lacks critical model infrastructure, i.e., an interface for image inputs or a decision algorithm.

The current implementation of the LCM model is limited by the fact that it cannot, at present, account for individual differences in reading abilities, reading experience, or knowledge of words. These limitations, however, are not principled limitations but a pragmatic decision to demonstrate the general feasibility and predictive value of the proposed lexical categorization process. Fitting the model to actual participants’ performance will require estimating parameters related to lexicon size, as these are the basis for deriving the word similarity parameter, which in turn is the basis for the lexical categorization process. The present results may also have implications for understanding dyslexia, i.e., a developmental deficit in reading performance. People who have dyslexia are characterized by altered structure and function in the subregion of lvOT that we related in the present work to the process of lexical categorization [5,6]. Consequently, novel hypotheses concerning the nature of the deficit(s) underlying dyslexia might emerge from LCM-based studies of the process of lexical categorization or the word-similarity signal on which the lexical categorization process operates. Extending the LCM model to fit individual subjects’ behavioral performance patterns would be interesting to test such hypotheses (i.e., as implemented here [51]). Also, modeling subject-specific behavioral performance would provide a novel approach to quantitative, model-driven diagnostics from which personalized training schemes could be derived.

Also, our current work was restricted to the German language. However, there are no principled reasons why the proposed lexical categorization process should apply more to (the orthography of) one language than others. Thus, future work should also explore the LCM model in other languages and orthographies (e.g., [51] for a potential use case). Lastly, the present work may also potentially be driven by the specific nature of pseudowords and non-words used in our work. However, during stimulus construction, we aimed to represent a wide variety of letter strings (concerning their word-likeness) so that the LCM evaluations reported here should be broadly generalizable. On the other hand, future work could test various types of pseudo- or non-words that, e.g., differ on specific word characteristics like morpheme structure.

The quantitative model comparison method applied in the present work is a straightforward way to compare models that assume different representations or different cognitive processes. In the context of the present work, this can, for example, be used to explore which representations are the most plausible input into the proposed categorization process (cp. S4 Fig) or the effects of different lexicon sizes on this process (S5 and S6 Figs), which may be relevant for understanding reading development. Such progress is hardly possible when models are exclusively based on verbal-metaphorical descriptions of cognitive processes. All models reported here are publicly available as code so that researchers can use them to test alternative and possibly more refined models. A crucial final stage of this approach, in addition to demonstrating the fit between model and data, however, will be to demonstrate causality of the assumed processes. The training-based evaluation we have reported in the present study is one plausible approach for achieving this, as it demonstrates the effect of manipulating the assumed process on behavioral performance in the respective domain.

In conclusion, the lexical categorization model (LCM), which is inspired by hierarchical (i.e., from occipital to anterior temporal to frontal regions) models of ventral visual stream processing, is a simple computationally explicit model that reliably describes lvOT activation patterns during visual word recognition elicited by a wide range of different letter strings. The LCM succeeds in simulating published seminal (and seemingly contradictory) fMRI results and is supported by empirical evaluations using tailored stimulus sets that differentiate between alternative models. Our empirical fMRI data suggest that a neural representation of word-likeness is estimated in brain regions posterior to word-sensitive lvOT cortex. This word-likeness representation is then fed forward into a lexical categorization process, which we could localize with new empirical fMRI data to a subregion of lvOT. Finally, after non-words are ‘filtered out’ by this categorization process, familiar words are passed through to higher-level cognitive processes such as lexical access and the extraction of word meaning, which are postulated to occur in downstream (temporal and possibly inferior frontal) areas. This framework (schematically visualized in Fig 7) is a significant step towards brain-based computational accounts [52] of information processing during visual word recognition and reading. It thus may in the future serve as an empirical basis for new interventions to help slow readers.

Methods

Supporting information

Acknowledgments

We thank Anne Hoffmann, Jan Jürges, and Rebekka Tenderra for help with fMRI data acquisition and Benjamin Peters for help with the entropy formulation. In addition, we thank Sophia Haan for helpful comments on a previous version of the manuscript.