Recovering an intended word from a noisy spelling is a general word-retrieval problem relevant to language learning, spelling correction, and cross-script search. We study this problem in Arabic, where learners produce variable transcriptions in Arabic or Roman script that may preserve only part of the target form. Using ArabScribe, we evaluate four large language models (LLMs) on ranked intended-word recovery under three cue conditions: the noisy form, an English meaning gloss, and both. Combining form and meaning consistently outperforms either cue alone across models and scripts, with particularly strong benefits for distorted and Roman-script input. An expert-human evaluation shows the same benefit from glosses and places the strongest models near individual expert performance on the sampled items. Analyses of edit distance, learner proficiency, and target frequency show that semantic information improves recovery across difficulty conditions but does not eliminate sensitivity to form noise and lexical rarity. These findings establish gloss-grounded intended-word recovery as a useful framework for studying how LLMs integrate form and meaning under noisy, cross-script input.

LLMs perform strongly across NLP, but their ability to produce explicit grammatical analyses remains unclear. Arabic provides a challenging testbed due to its rich morphology and orthographic ambiguity, which create strong morphology–syntax interactions. We present a unified evaluation of LLMs on Arabic morphosyntactic tagging and dependency parsing, covering pre-tokenized, raw-text, and cascaded settings. We compare zero-shot prompting with retrieval-based in-context learning. Relevant demonstrations substantially improve performance. The strongest LLMs approach supervised tagging and parsing systems; however, they require substantial annotated data for demonstration retrieval and considerable computational resources. We make all code and data used in this paper publicly available.

We present BAREC-TB, the first dependency treebank of Arabic children’s literature and educational texts. The 156K-word corpus includes texts drawn from eight sources, including curricula, reading materials, children’s stories, poems, and songs. The treebank is annotated according to the Columbia Arabic Treebank (CATiB) dependency representation. We compare BAREC-TB with existing Arabic treebanks, including the Penn Arabic Treebank and Camel Treebank, using corpus statistics, dependency-based measures of syntactic complexity, readability and vocabulary-related features. We also establish parsing benchmarks using both a trainable parser and large language models. The corpus will be made publicly available to support Arabic NLP, educational NLP, and language pedagogy research: https://barec.camel-lab.com/.