Efficient reading is crucial for societal participation, and improving reading proficiency is a major educational goal. Slow reading, prevalent in individuals with dyslexia and migrant language learners, hinders access to information and negatively impacts socioeconomic status. This study focuses on non-native German speakers to evaluate a support program designed to improve reading skills. The program is grounded in a neuro-cognitive computational model, the Lexical Categorization Model (LCM), which describes pre-lexical orthographic processing implemented in the left-ventral occipital cortex. The LCM posits that this area performs lexical categorization, distinguishing words from non-words. The study combines LCM-based training with individualized machine learning diagnostics to create an effective and personalized training procedure. The machine learning component allows for the identification of key processes predictive of training success. Fast and efficient visual word recognition is central to reading, transforming visual information into meaning. Less proficient readers rely more heavily on this process for comprehension. This study evaluates a training procedure aimed at improving word recognition to increase reading speed, drawing parallels to successful phonics training programs which are developed from well-evaluated cognitive models, like dual-route models, to train grapheme-phoneme associations. The research incorporates brain imaging findings which show a strong link between the visual word form area (in the left occipito-temporal cortex) and efficient word recognition, with reduced activity observed in slow readers and illiterates. The LCM, distinguished from other models by its lack of free parameters, provides a transparent framework for designing a targeted training program and developing a highly transparent prediction model for individualized diagnostics.

The literature extensively supports the link between efficient visual word recognition and overall reading performance, impacting typical readers, dyslexics, language learners (L2), and beginning readers. Several studies demonstrate that improving visual word recognition is critical for comprehension in less proficient readers. This study builds upon previous research using model-based training approaches, such as phonics training, which has shown small to medium effect sizes on improving reading skills in young dyslexic readers by targeting grapheme-phoneme associations. Neuroimaging studies have consistently shown activation in the visual word form area (vOT) during word recognition, with reduced activity in slow readers and illiterates. While several theoretical models attempt to explain vOT activation, the LCM stands out due to its computational implementation and transparent features. The LCM’s core assumption is that the left-ventral occipito-temporal cortex implements a lexical categorization process to filter non-words from further processing. This categorization relies on word-likeness, with uncertainty highest when word and non-word distributions overlap. Previous research has shown that the LCM accurately models the activation patterns in the left-ventral occipito-temporal cortex, exceeding the performance of alternative models.

Three experiments were conducted with 76 adult non-native German language learners (after excluding participants due to procedural issues, erroneous participation, or technical reasons). Experiment 1 served as a pilot study for the lexical categorization training, while Experiments 2 and 3 employed randomized controlled trials to compare the lexical categorization training with alternative approaches. The core of the lexical categorization training involved a lexical decision task (word or non-word) with feedback, aiming to improve the lexical categorization process as described by the LCM. Experiment 2 compared the lexical categorization training with phonics training (grapheme-phoneme association), and Experiment 3 compared it with a modified lexical categorization training that involved changing fonts to train the formation of orthographic representations. Reading speed was measured using the adult version of the Salzburger Lesescreening (SLS), a paper-and-pencil test assessing sentence-level reading comprehension, administered before and after the training. Individualized machine learning diagnostics were implemented to predict training outcome using LCM parameters (lexical categorization uncertainty), other word characteristics (orthographic Levenshtein Distance 20, or OLD20), incoming reading speed (SLS), and response times and accuracy from the first training session. A leave-one-out cross-validation procedure was employed to prevent overfitting, with consensus-nested leave-one-out cross-validation used for hyperparameter tuning (feature selection and model type). Three model types were evaluated: multiple regression, support vector machine (with linear and radial kernels), and random forest. Feature importance was assessed by the frequency of feature selection across the cross-validation runs and t-values from the best-performing model.

In all three experiments, the lexical categorization training resulted in a significant improvement in reading speed (Experiment 1: 26.6%; Experiment 2: 20.4%; Experiment 3: 22.4%). Response times during training consistently showed lexical categorization uncertainty effects (higher reaction times for hard-to-categorize items). The phonics training in Experiment 2 did not significantly improve reading speed but showed a strong learning effect on response times, suggesting that the task itself influences reading speed, rather than just stimuli. The modified lexical categorization training (font changes) in Experiment 3 resulted in a significant reading speed increase, but not significantly higher than the standard lexical categorization training. The machine learning pipeline successfully predicted the outcome of the lexical categorization training, with a correlation between predicted and observed training effects of 0.69 (p<0.001). The best-performing pipeline used multiple regression, a feature selection criterion of 10, and a predictor composition including a three-way interaction of lexical categorization uncertainty, OLD20, and sequence index. Feature importance analysis indicated that LCM-related features (lexical categorization uncertainty, lexicality, and OLD20) were frequently selected along with incoming reading speed and training week. Applying the prediction model to select responders increased the mean reading speed improvement from 23% to 43%.

This study's findings support the central role of lexical categorization in efficient reading. The consistent increase in reading speed across three experiments, coupled with the high feature importance of LCM parameters in predicting training success, provides strong evidence for the direct association between lexical categorization and reading efficiency. The null effect of phonics training in Experiment 2 might be attributed to the participant group (fluent first-language speakers with limited German knowledge), suggesting that the effectiveness of phonics training may differ across populations. The higher training effect observed in the modified lexical categorization training (Experiment 3) suggests the possibility of training visual predictive processes, though further replication is needed. The successful prediction of training outcomes using machine learning, relying heavily on LCM parameters, demonstrates the potential for individualized diagnostics to optimize resource allocation in reading interventions. The results highlight the synergy between neuro-cognitive computational models, targeted training programs, and individualized machine learning diagnostics in improving reading skills.

This study demonstrates the effectiveness of a lexical categorization training program for improving reading skills in non-native German speakers. The findings underscore the importance of lexical categorization in efficient reading and showcase a novel framework for investigating visual word recognition processes. The combination of a transparent computational model (LCM), a targeted training procedure, and explainable machine learning diagnostics offers a powerful approach for developing individualized reading interventions. Future research should focus on expanding this framework to include other training procedures, populations (e.g., dyslexics), and exploring more sophisticated machine learning techniques while addressing the limitations of sample size and the need for independent sample validation.

The study's reliance on cross-validation necessitates further validation with an independent sample to fully address potential overfitting. The sample size, though powered for Experiment 3, may limit the generalizability of the findings. The reliance on self-report measures of motivation may not fully capture the level of participant engagement. Furthermore, using an alternative measure of reading proficiency, such as eye-tracking, may provide a more comprehensive assessment of reading skills. Finally, the study did not explicitly investigate the individual mechanisms underlying reading improvement. Therefore, more research is needed to fully elucidate the causal links and gain a better understanding of the cognitive mechanisms underlying improvements in reading.