Why Current XAI Is Not Enough for Arabic NLP: A Critical Survey of the Explainability Gap

Explainable AI (XAI) is now a major theme in NLP; however, Arabic NLP remains under-explained in three connected senses. First, there is a method gap: Arabic XAI relies heavily on a small set of post-hoc techniques such as LIME, SHAP, attention visualization, and saliency, while broader NLP XAI offers richer diagnostic, counterfactual, probing, rationale-based, and human-centered methods. Second, there is a task gap: existing Arabic XAI work is concentrated in classification tasks, especially sentiment analysis, hate/offensive language detection, fake news, and spam, with weaker coverage of generation, retrieval, translation, summarization, structured prediction, and dialogue. Third, there is a linguistic gap: many explanations identify influential tokens, but rarely explain Arabic-specific phenomena such as morphology, clitics, dialectal variation, diglossia, orthographic ambiguity, diacritics, code-switching, named entities, cultural references, or Classical and religious registers. This critical structured survey synthesizes the reviewed literature on Arabic XAI across text, speech, and multimodal settings. We argue that Arabic NLP does not only need explanations of model decisions; it needs explanations that are faithful to Arabic as a linguistic, cultural, and sociotechnical object. We introduce a taxonomy of tasks, methods, linguistic units, varieties, goals, and evaluation practices, and propose a research agenda for linguistically grounded Arabic XAI.

Self-Training Doesn't Flatten Language -- It Restructures It: Surface Markers Amplify While Deep Syntax Dies

Successive self-training on a language model's own outputs is widely characterized as a process of flattening: diversity drops, distributions narrow, and the text becomes "more like itself." We provide evidence that this characterization is incomplete. Across eleven generations of self-training on five models (GPT-2 124M, Pythia-410M, Pythia-1.4B, OPT-1.3B, Pythia-2.8B), language is not flattened uniformly -- it is restructured. Surface markers (discourse connectives, hedges, em-dashes) rise, while mid- and deep-syntactic structures (questions, parentheticals, passives, subjunctives) collapse. We formalize this asymmetric collapse as the Structural Depth Hypothesis (SDH): the per-generation decay rate of a linguistic feature is predicted primarily by its structural depth -- the number of nested syntactic dependencies it requires -- and only secondarily by its generation-zero output frequency. Pooling 17-feature panels from five models spanning three architecture families (N=85), the pooled Spearman correlation is rho=0.540 (p < 10^{-6}; cluster-bootstrap 95% CI [0.434, 0.634]), while frequency is a substantially weaker predictor (rho=0.225). A matched human-text fine-tuning control yields rho=0.039 (p=0.88), confirming the gradient is self-training-specific. We further document a Superficial Complexity Paradox: aggregate complexity proxies (dep-tree depth, TTR, word length) all rise as the underlying clause structure dies, with direct implications for training-data curation and LLM-text detection.