The past two decades has seen a resurgence of interest in the idea that comprehenders routinely engage in prediction of upcoming linguistic content. Prediction is now often assumed to occur along all levels of linguistic representation, from semantics and morphosyntax to phonology/orthography, serving to facilitate the access and integration of bottom-up information into unfolding sentential and discourse representations.

A key observation supporting theories that comprehenders predict upcoming words comes from experimental research that exploits morphosyntactic and phonotactic constraints on the form of words that precede high cloze target words. DeLong, Urbach, and Kutas (2005) uses the phonotactic constraints of English indefinite articles to investigate probabilistic pre-activation of expected (high-cloze) nouns. They presented participants with highly constraining contexts like “The day was breezy so the boy went outside to fly…” and manipulated the article and noun so that the noun was either expected, a kite, or unexpected, an airplane. The amplitude of the N400 elicited by the preceding a/an articles was found to decrease as the cloze probability of nouns increased. Other ERP studies investigating morphosyntactic constraints (Foucart, et al., 2014; Martin, et al., 2018; Otten & Van Berkum, 2009; Wicha, Moreno, & Kutas, 2004), phonotactic constraints (Martin, et al., 2013), or both (Ito, et al., 2020) report evidence for pre-activation of target nouns on preceding articles. Although some studies have not successfully replicated these effects (Ito, Martin, & Nieuwland, 2017; Nieuwland, Arkhipova, & Rodríguez-Gómez, 2020; Nieuwland, et al., 2018), taken together this research has underpinned theories of prediction in language comprehension by providing unambiguous evidence for predictive processes.

Critical to the success of some of these studies is the high temporal resolution and word-by-word presentation typical of ERP methodology in reading. This is particularly true for phonotactic/orthographic constraints which occur between adjacent words. These tight sequential constraints prevent researchers from investigating these phenomena with traditional incremental reading tasks. Self-paced reading is one of the most commonly used techniques and is relatively cheap, especially with recent innovations in online participant crowdsourcing and web-based stimulus delivery. However, it is difficult to isolate effects on a specific word due to well-known spillover effects which smear processing differences across multiple words (Mitchell, 1984; Witzel, Witzel, & Forster, 2012). Eye movements in reading are also a poor measure. Short functional words are often skipped during reading and are likely available in parafoval preview, requiring very careful experimental manipulation to isolate effects on them (Balota, Pollatsek, & Rayner, 1985; Cutter, Martin, & Sturt, 2020).

Research on prediction has been held back by the lack of a cheap and easy behavioral methodology to study early cues to prediction error (cf. Van Berkum, et al., 2005). The maze task, however, offers an alternative incremental reading method that can address some of the potential shortcomings of the methods mentioned above (Forster, Guerrera, & Elliot, 2009). In maze tasks, sentences are presented to participants as a sequence of choices between two alternatives. One alternative is the correct continuation of the sentence while the other is a distractor. Distractors are real words that are anomalous given current sentence context (G(rammaticality)-maze) or pseudowords (L(exicality)-maze). Maze tasks have been shown to deliver focal measures of processing difficulty, with effects occurring on their trigger word, that are comparable to self-paced reading and eye tracking (Witzel, Witzel, & Forster, 2012; Witzel & Forster, 2014).

Maze tasks have been difficult to implement because pairing a distractor word to each word in all sentence material was labor-intensive and prone to researcher error. Addressing these difficulties, Boyce, Futrell, and Levy (2020) used natural language processing to automate distractor selection, creating the A(uto)-maze. They demonstrated that the A-maze elicited focal reading time effects on the disambiguating word of three types of attachment ambiguities with online participants. A-Maze effects were comparable to G-maze and more sensitive than L-maze and self-paced reading.

Given that focal effects can be found with A-maze, A-maze response times might also isolate prediction error effects on expectation-mismatching article forms preceding unexpected nouns. As a participant-controlled continuous temporal measure, A-maze response times might also provide new information about predictive processing, e.g. via natural differences in participant’s comprehension speed. To investigate these possibilities, an A-maze task was conducted on high cloze probability sentence contexts manipulating noun predictability and the form their immediately preceding articles.

By providing a focal reading time measure, the maze task was able to reveal effects of expectation both on target nouns and the article a/an contrast preceding the nouns. Unexpected nouns and their preceding articles, which mismatched the expected noun’s required article form, were responded to more slowly than expected nouns and their preceding articles. Response times on articles and nouns were also found to be inversely related to noun cloze probabilities, with response times decreasing as cloze probabilities increased. This suggests that pre-activation of word form is a more graded effect (DeLong, Urbach, & Kutas, 2005).

Interesting, early predictive effects of article form were magnified for slower responders and smaller with faster responders. This suggests that probabilistic pre-activation of the expected word to the level of phonological form that is required to compute expectations for preceding article form may take time to emerge, consistent with prediction-as-production theories (Pickering & Gambi, 2018). The maze task may be sensitive to these effects because its response times are generally longer than self-paced reading’s reading times, eyetracking’s fixation durations, or ERP’s typical SOAs. This additional time may have aided form prediction. Ito, et al., (2016), for example, found that N400-reduction to form-related words emerged at 700 msec SOA but not 500 msec SOA. A similar effect may have emerged in a more natural and graded fashion here through the variation of individual’s average response times. Slower responders may have given themselves more time to reach a form prediction for the upcoming noun and compute its consequences for the preceding article’s form, while faster responders were less likely to make these form predictions and consequent computations.

These results demonstrate that the maze task can be sensitive to the predictive use of phonotactic constraints between an expected word and its preceding word, and may be a useful alternative methodology for investigating predictive comprehension mechanisms. A-maze eases the burden of distractor generation and is effective with online crowdsourced participants (Boyce, Futrell, & Levy, 2020). It is hoped that further investigation into predictive mechanisms will be spurred on with this demonstration of the maze task’s effectiveness.