The prediction-enhanced lexical processing where parser incrementally incorporates parsed lexical items to predict upcoming words is strengthened by anticipatory evidence especially when constrained by context across measures (N400, cloze probability, surprisal).
 
But what about prediction-enhanced morphological decomposition?, the parsing model where the parser uses contextually determined lexical predictions to guide how words are broken down into their constituent morphemes — roots, prefixes, suffixes, and inflections.
 
The core idea: instead of applying rule-based or statistical segmentation blindly, the model first generates a prediction about what the next item’s morphological-representation likely is given its context, then uses that expectation to build the most plausible morphological parse or representation. This matters because prediction enhanced morphological representations can facilitate decomposition to the root even when the root is not lexically matched to the contextually predicted lexeme but is to the morphological-representation of the upcoming noun.
 
Individuation-ambiguous roots such as family often have multiple plausible representations, and the correct parse changes the item's semantics substantially. The key insight is the feedback loop between prediction and decomposition. A pure morphological parser treats family as a sequence of morphemes in isolation. A prediction-enhanced parser knows from context that the word is being used as a individuated-mass noun as in “my family is numerous” or individuated-count noun as in “my family are numerous”, which in the latter case strengthens the parse “my family.PL” and supresses less compositional representation.
 
A model enhanced by morphological prediction can use predictive context to facilitate decomposition in languages with rich morphological representations and benefits substantially, since the prediction prior compensates for root-ambiguity.
 
The model is tested in Croatian using individuation-ambiguous (mass vs count) nouns in a self-paced reading and lexical decision task.

In recent decades, diverse frameworks in linguistic theory have come to the consensus that there is no common unit of representation that unites linguistic form, syntax, and meaning corresponding to the colloquial 'word' (e.g., Haspelmath 2011). However, most  psycholinguistic models and cognitive neuroscience of language models continue to assume that language processing is organized around such 1-to-1-to-1 mappings, whether they call them 'lexical items', 'morphemes', or 'lemmas'. This kind of approach discounts the great prevalence of so-called 'idiomatic' non 1-to-1 mappings across human languages, as well as the fact that phonology, like syntax and semantics, has a rich dependency structure in its own right that is not isomorphic to either (Krauska & Lau 2023). But even for psycholinguists who recognize this, it can be hard to resist adopting the well-trodden 1-to-1 framework when they just want to get on with studying their phenomenon of interest. In this talk I'll use a recent study from our group investigating the production of English compounds with EEG (Cuonzo et al. 2026) to illustrate how bread-and-butter language processing research can be designed and presented in a way that is more consistent with the known properties of natural (spoken or signed) human languages.