Confronting the Interaction Problem in Arabic NLP: A Resolve Layer Framework

Arabic NLP continues to exhibit persistent failure modes that cannot be explained by morphological complexity or dialectal variation alone. While existing systems address these challenges through specialized components such as dialect identification, normalization, and diacritization, they treat each dimension of variation as separable. In practice, much Arabic input reflects intertwined patterns of dialect, register, and orthographic convention, often improvised and inconsistently blended, which resist uniform processing. Errors persist when systems commit to fixed methods without dynamically adapting in response to shifts in the input. This paper names the interaction problem and proposes the Resolve Layer, a framework treating adaptive regime identification as a first-class processing function. The core architectural claim is that resolving interpretation belongs at intake, immediately after tokenization, not downstream from normalization and diacritization. Regime-conditioned interpretation operates at the phrase and construction level, not just individual tokens, enabling the system to resolve identical strings differently depending on dialect, register, and context. As a position paper, it articulates the problem, specifies what an adequate response would require, and proposes an architectural path for empirical exploration.