📄 Arabic Diacritic Recovery Using a Feature-Rich biLSTM Model

Summary

Arabic is usually written without short vowels (diacritics), so readers must supply them to pronounce words and resolve meaning. The authors split the task in two. Core-word (CW) diacritics determine which lexical item is meant. Case endings (CE) on the stem's final letter signal syntactic role and depend on context that may be far away. Each part gets its own biLSTM. A character-level model handles CW diacritics, guided by word segmentation and a unigram language model, and the same unigram model also corrects its output afterwards. A word-level model predicts CE from a broad set of surface, morphological and syntactic features, including POS tags, gender, number, morphological templates and affixes. Training uses a 9.7M-token diacritized MSA corpus (the one behind RDI and Farasa) and a 2.7M-token Classical Arabic (CA) collection. MSA is tested on the WikiNews set. Full-word error falls to 6.0% on MSA, about half the previous best of 12.2%, and reaches 4.3% on CA. The results show that deliberate feature engineering still pays off inside neural sequence models.

Key points

• MSA results: CWER 2.86% and CEER 3.7%. For CE, Farasa scores 10.7% and RDI 14.4% on the same data, so the new error is roughly 60% lower.
• CA results: CWER 2.2% and CEER 2.5%. Full-word WER (CW+CE) is 6.0% for MSA and 4.3% for CA.
• Data: the MSA training corpus has 9.7M tokens and about 194K unique surface forms. The WikiNews test set has 70 articles (18,300 tokens). For CA, about 400k words (5,000 sentences) of a publisher's 2.7M-token collection were held out for testing.
• Running the best MSA model on CA raises CEER to 8.9%, so each variety needs its own training data. In the CA ablation, POS features help most, followed by surface features.
• The most frequent CE error is swapping fatHa and damma (19.3% of errors). The main causes are POS mistakes and subject/object confusion.

Interesting ideas

• I1. The lexical part of the problem (word-internal vowels) and the syntactic part (case endings) get separate models, one per character and one per word, so each sees the right context. — For Mova (tokens): Mova could keep lemma and stem prediction apart from context-dependent inflectional features, assigning each to a component whose window and input units suit it.
• I2. A plain unigram lexicon both informs and post-corrects a neural character model, pulling odd outputs back toward forms actually seen in the data. — For Mova (tokens): The lemmatizer and morphology output could be checked against a frequency lexicon, with a logged and explainable fallback whenever the model proposes an unseen form.
• I3. Hand-crafted linguistic features (POS, gender, number, templates, affixes) are fed into a biLSTM and then ablated one by one to measure what each contributes. — For Mova (tree): gparse could take explicit morphological features alongside raw tokens. A feature-by-feature ablation would show which cues the parser actually relies on.
• I4. Errors are grouped by their underlying cause and split into those fixable with better upstream tools (POS tagging, name detection) and those that need deeper understanding (attachment, subject/object). — For Mova (learn): Evaluation reports could trace each error back to the pipeline stage responsible, which would show whether a fix belongs in the tagger, the parser or the frames.
• I5. The authors stress applying one annotation convention consistently, note errors in their own reference data, and plan a test set that accepts several valid answers. — For Mova (learn): Silver data and UD dialect conversions should be checked for convention drift. Questions with more than one valid analysis could be scored against multiple references rather than a single gold label.