📄 Morphological Segmentation and Part of Speech Tagging for Religious Arabic

Summary

Most Arabic NLP tools are trained on newswire Modern Standard Arabic. Pre-modern religious writing, which is closer to Classical Arabic, uses vocabulary and constructions those tools rarely see. The author built a small hand-checked corpus from three religious booklets: Nawawy's prophetic traditions, the legal primer Matn, and Gazaly's philosophical Munqith. The texts were first processed automatically and then corrected by hand, giving about 27K words with fine-grained, ATB-style segment boundaries and segment-level POS tags.

Memory-based learners (Timbl for segmentation, MBT for tagging) were trained three ways: on the ATB, on the religious data, and on both combined. Despite being 21 times smaller than the ATB, the religious data won clearly:

• segmentation rose from 90.84% to 95.70%;
• POS tagging rose from 82.22% to 91.26% with gold segmentation;
• POS tagging rose from 78.62% to 88.22% with automatic segmentation.

Adding the ATB to the religious data helped only marginally. A coverage check against Gigaword also showed many unseen words. The author concludes that more newswire data will not close the gap and that religious Arabic needs its own treebank.

Key points

• The corpus has three books totalling 27,442 words, 8,686 types and 47,054 segments, split once into fixed random 80/20 train/test sets by sentence.
• Segmentation word accuracy was 90.84% for ATB-trained, 95.17% for Religious-trained and 95.70% for combined training. OOV dropped from 44.80% to 23.3%.
• Whole-word tagging with composite tags reached 78.44% (ATBWW), 85.90% (RELWW) and 86.96% (combined). Repeating the religious training data 10 times raised this to 87.24%.
• The ATB training set contains 139 segment tags, and the religious set contains 117. Six tags appear only in the religious set, three of them imperative-verb suffixes. ADJ, PV and NOUN_PROP were the hardest tags under both training regimes.
• On a 148M-word slice of the Arabic Gigaword, religious vocabulary still showed 22.82% OOV by type and 9.32% by token.

Interesting ideas

• I1. The author annotated at the finest granularity possible, marking every inflectional and clitic boundary, because coarser segmentations and tagsets can be derived from it mechanically later.
• I2. Assimilated forms were annotated as they are actually written instead of restoring underlying morphemes, since the original information is easy to recover once segments are tagged correctly.
• I3. Upweighting the small in-domain set by repeating it many times inside a concatenation with a large out-of-domain set gave the best whole-word tagger.
• I4. The author checked OOV coverage against a huge general corpus to argue that more out-of-domain data would not fix the gap.
• I5. The evaluation was strict: a word counts as correct only if all its boundaries, or its full composite tag, are right, and results were reported with both gold and automatic segmentation.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (tokens): Storing maximally split morphology in crate en would let different downstream consumers project their own coarser tag views without re-annotating.
• I2 (tokens): English contractions and fused forms could be kept surface-faithful in the tokenizer, with a deterministic mapping to canonical lemmas recorded as an explainable step.
• I3 (learn): When mixing small genre-specific silver data, such as fables, with large general treebanks, oversampling the in-domain part is a cheap baseline to test before more complex adaptation.
• I4 (learn): Measuring type- and token-level OOV of a target genre against our training material before collecting data tells us whether to scale general data or build in-domain resources.
• I5 (tokens): Reporting tagger accuracy both on gold tokens and on our own tokenizer output would isolate how much error propagates from tokenization.