Prediction plays a critical role in comprehending human language. Many theoretical and computational models have attempted to characterize how we use context to facilitate language processing in noisy environments either with or without relying on predictive processing. Despite these attempts, we do not yet have a complete understanding of the role of prediction in language processing. Predictive coding models have recently gained popularity as potential architectures for the role of prediction during language comprehension. These models suggest that predictions about bottom-up inputs are continuously generated from higher cortical levels to lower levels in a hierarchical manner – i.e., a particular level generates predictions about the next lower level. As bottom-up input is encounter by each level of processing, prediction error is computed by comparing the input with the top-down prediction. The goal of this dissertation was to assess whether predictive coding models can account for the time course of information retrieval during predictive language processing. Specifically, the studies described examine the time course of pre-activations of lexical and sub-lexical features in both monolinguals and bilinguals using a combination of decoding electroencephalogram (EEG) with machine-learning classifiers and mass univariate event-related potential analysis. Chapter 1 describes an experiment that compared three frequently used models for signal classification – support vector machines (SVM), linear discriminant analysis (LDA), and random forest (RF) to determine which is best-suited for analyzing word pair prediction paradigms. Results showed that SVM was the best performing classifier of the three within two data sets from separate visual word priming paradigms. Chapter 2 describes an experiment which used EEG decoding with SVM classifiers and mass univariate ERP analyses to identify the time course of information retrieval prior to the onset of accurately predicted, related but inaccurately predicted, and unrelated target words during a visual word priming prediction paradigm. In addition to this pre-stimulus information retrieval, these analyses were used to investigate the effects of prediction error. The results of this study showed that semantic information, such as concreteness, is retrieved earlier than visual feature information, like word length, and that unrelated words had greater prediction error than predicted or related but inaccurately predicted words. Finally, Chapter 3 describes an experiment that extends the results of Chapter 2 by using the same paradigm and analyses with Spanish-English bilingual participants. The results of this study showed that bilinguals reading words in their second language (L2) retrieve anticipated information in a similar fashion as monolinguals. Semantic information preceded visual information and unrelated words showed evidence of greater prediction error than did predicted or related words that were not accurately predicted. Together, these experiments support predictive coding models of language processing in both monolinguals and bilinguals during word recognition. Both groups predict higher-level features (concreteness) before lower-level features (word length) of anticipated words and calculate prediction error when they make inaccurate predictions.

Language processing is highly complex, yet the brain performs this task effortlessly. One explanation is that the brain continuously generates top-down predictions about upcoming word based on prior context, reducing the need for further bottom-up processing. Previous research has shown that the processing of predictable words in constraining contexts is facilitated, and that the semantics of a predicted word can be pre-activated. However, it remains unclear which other types of representations may be pre-activated. Furthermore, theories have demonstrated the role of the production system in generating prediction, which has not been directly tested. Therefore, this dissertation aimed to systematically examine the levels of representation that may be pre-activated by lexical predictions (i.e., semantic concreteness, orthographic and phonological neighborhood size, word length), as well as the impact of prior predictions generated by production on subsequent visual word processing.

Chapter 1 demonstrated an experiment that used a semantic priming paradigm in conjunction with a silent prediction task to assess the influence of prediction accuracy on lexical processing of the target word. Participants read 480 sets of prime and target word pairs, while silently predicting the target after reading either a semantically related (circus - CLOWN) or an unrelated prime (trim - CLOWN), followed by a lexical decision task on the target. There were 125 filler sets of word-nonword pairs. We compared the target lexical decision RTs based on prediction accuracy and semantic relatedness. Results showed significant effects of prediction accuracy and typical semantic priming effects, as well as pre-activation of orthographic and phonological word form. Thus, when no overt production is required, participants pre-activated the word form level feature that was the most relevant to meet the task demand.

To directly assess the precise status and content of prediction, Chapter 2 replaced the silent prediction task with a self-generated prediction task, ensuring the predicted word is fully activated. Participants were asked to speak the first word that came into mind before seeing the target, followed by a lexical decision on the target. The lexical decision RTs were compared across three conditions based on the predictive naming accuracy and the target-response relatedness: same-related (prime circus – produced clown – target CLOWN), different-related (prime circus – produced ELEPHANT – target CLOWN), different-unrelated (prime trim– produced CUT – target CLOWN), Results showed main effects of naming accuracy and semantic relatedness, and pre-activation of phonological form.

Chapter 3 used ERPs to investigate the influence of prior predictions made by the production system on subsequent visual word processing. In other words, instead of processing the word in a bottom-up fashion, the process resembled an error signal for revision, followed by the mapping of visual information and then semantic processing/updating. We assessed the main effects on the early Error-related Negativity (ERN) component to examine the prediction error detection, followed by the P2 component for the mapping of visual input, and then, the N400 for semantic processing and updating. Results showed early effects of prediction accuracy on ERN amplitudes elicited by different-related trials, indicating an error signal based on visual input. This signal was only triggered when the new visual input was associated with the predicted word form, allowing it to prompt learning and improve future prediction accuracy. During the P2 window, we found main effects of naming accuracy and semantic relatedness, suggesting that the effect of relatedness delayed prediction by approximately 100ms. Also, we found effects of phonological neighborhood size on the P2 in only the same-related condition, reflecting mapping of expected information. Also, effects of orthographic neighborhood size and word length on the P2 amplitudes were found in the different-related trials, suggesting that the brain continued to use visual input to elicit learning signals for the next processing stages. Finally, during the N400 window, the main effects remained significant. We also found significant effects of orthographic neighborhood size on the N400 amplitudes elicited by the different-related and different-unrelated trials, but not in the same-related trials, indicating that visual processing may have completed for accurately predicted words. However, the semantic concreteness effect on the N400 was significant across all three conditions, regardless of prediction accuracy. We argue that for differently named words, the N400 effect may indicate further revision of representations, whereas for same-related words, semantic processing is largely complete.