The reading procedure of the BRAID-Acq model is best described as "analogy-based" because it relies exclusively on lexical knowledge, without incorporating abstract knowledge of the spelling-sound relationship.

Hypothesis

We hypothesize that both known and novel words engage the same components of the model, contrasting with the theoretical framework of dual-route models (Pritchard et al., 2018; Ziegler et al., 2014). Specifically, we suggest that both known and novel words can be read using the same set of lexical knowledge. Additionally, we explore the hypothesis that segmenting a novel word into graphemes is unnecessary for accurate reading; instead, visuo-attentional fixations could allow for processing the novel word in smaller parts. Thus, we propose that processing novel words does not depend on any fixed psycholinguistic unit.

Description of the mechanism

In the BRAID-Acq model, the reading procedure is both lexical, meaning it uses lexical knowledge, and sub-lexical, meaning it involves processing the stimulus into smaller parts. Reading a novel word involves several successive fixations. During each fixation, the model selects words that are orthographically similar to the stimulus at the portion currently being processed. The model then uses the phonological representations of these selected words to calculate a likely pronunciation.

During incidental learning, readers encounter both words with and without prior phonological knowledge. Accurately learning both types of words without knowing if their pronunciation is familiar beforehand is a challenge for both readers and computational models. For example, when encountering "debt," a reader is likely to decode it as /dɛbt/, while the actual pronunciation is /dɛt/. In attempting to match the pronunciation with a phonologically familiar word, the reader might associate the orthographically novel word with "det." But now, imagine a novice English reader decoding the pseudo-word "shep" as /SEp/. The reader must decide whether to associate it with an existing phonological form like /SEIp/ ("shape") or treat it as completely novel, necessitating the creation of new orthographic and phonological representations for "shep" and /SEp/. This decision step poses a significant challenge for computational models because there is no easy way to differentiate between the two situations: in both cases, the pronunciation matches no entry in the lexicon. Therefore, how can one learn phonologically new words while remaining flexible enough to accommodate words with prior phonological knowledge?

Hypothesis

Unlike the choices made by two computational self-teaching models (Pritchard et al., 2018; Ziegler et al., 2014), and in accordance with behavioral data (Nation & Cocksey, 2009), we propose that knowledge of a word's phonological form and the presence of context are not essential for learning. However, these elements are essential to be able to manage the variety of incidental learning situations.

We believe context to be the most efficient way to handle the fact that learning involves learning both words with and without prior phonological knowledge.

Hypothesis

In accordance with the self-learning theory, we hypothesize that context should: 1) guide reading towards a set of plausible words without favoring a specific lexical item, and 2) suggest rather than constrain, meaning that learning possibilities are not limited to words suggested by contextual cues.

Description of the mechanism

Contrary to dual-route models of reading acquisition, in the BRAID-Acq model, context plays a role throughout the entire processing phase. During perceptual processing, context has two effects. First, the model prioritizes the identification of words belonging to the context. Second, it impacts phoneme identification, a process we could term "pronunciation correction." This mechanism is used while reading words with and without prior phonological knowledge. Only words that are both 1) part of the semantic context and 2) whose phonological representations closely match the phonemes identified during decoding, contribute to this pronunciation correction. Ultimately, when the stimulus belongs to the context and decoding is only partially incorrect, then an online mechanism gradually corrects decoding errors, resulting in accurate pronunciation.

Real-life situations also involve reading without meaningful context. The model BRAID-Acq also implements a mechanism to support novel word learning in this case.

Hypothesis

Phonological lexical knowledge supports the model's reading process in a "top-down" manner (from lexicon to phonemes).

Description of the mechanism

The model includes a top-down lexical mechanism modulated by phonological lexical familiarity: the more certain the model is that the stimulus belongs to the phonological lexicon, the more phonological lexical knowledge supports phoneme identification. Specifically, when the model "hesitates" between two phonemes, if the stimulus is evaluated as sufficiently familiar to trigger top-down lexical processing, this ultimately enables the model to make a correct decision.

Simulations with the BRAID-Acq model have demonstrated the ability to read novel words without relying on graphemic segmentation and grapheme-phoneme conversions. Even in the absence of context, top-down lexical feedback allows the model to correct some pronunciation errors. Moreover, when context is present, it significantly improves the model's accuracy in reading words with prior phonological knowledge, while maintaining pronunciation accuracy for words without prior phonological associations. Thus, the model effectively handles a variety of situations!