Toward Cognitively Grounded Machine Translation: The Morpho-Priming-Wait Model for Lexical Ambiguity in English-Arabic Neural Translation

Contemporary neural and large-language-model translation resolves lexical ambiguity implicitly: whatever disambiguation occurs is distributed across learned representations, with no stage at which a reading is chosen, recorded, or can be corrected. This position paper argues that ambiguity resolution should instead be an explicit, cognitively modelled, and inspectable stage of the translation pipeline, mirroring the sequence human comprehenders follow—detect that an item is uncertain, withhold commitment, let context pre-activate the licensed reading, and only then choose. We specify the Morpho-Priming-Wait (MPW) model, which couples three mechanisms drawn from human sentence processing. Connectionist morphology grounds lexical representation in Arabic root-and-pattern structure, keeping competing analyses of a written form distinguishable rather than resolving them in advance. Priming realizes spreading activation computationally, letting context raise the accessibility of the reading it licenses before a choice is made. A wait-and-see discipline defers commitment where either the output or the analysis is uncertain, revising it once further evidence accumulates. English-Arabic translation supplies unusually direct motivation, because Arabic orthography marks neither short vowels nor the boundary between a proclitic and its host: a user seeking either reading types the identical string بطاقة which may be the noun biţaqah 'card' or the preposition bi- together with the noun ţaqah 'capacity'. The system must therefore recover how many words it is reading before it can decide what they mean, a decision no sense inventory can express. We situate each mechanism against self-attention, sense-aware neural translation, and confidence-gated iterative decoding, locating MPW's contribution in their coupling and in the legibility of the decisions they reach. Because the proposal is theoretical, we close by specifying a disambiguation-accuracy benchmark, the ablations that would isolate each component, and the conditions under which the proposal should be rejected.