📄 Book Reviews: Arabic Computational Morphology: Knowledge-Based and Empirical Methods by Abdelhadi Soudi, Antal van den Bosch, and Günter Neumann (editors)

Summary

George Kiraz reviews an edited Springer volume on Arabic computational morphology (Text, Speech, and Language Technology series, vol. 38, 2007). He frames it against the early 1990s, when the field had many proposals but few implementations, with work by Beesley and Buckwalter as the main exceptions. After a preface by Richard Sproat, the fifteen chapters fall into four parts:

• Introductions: a transliteration scheme, plus orthography, encoding history, tokenization and dialect issues.
• Knowledge-based methods: syllable-based and inheritance-based accounts in DATR, the Lisp-based MORPHĒ system, and a chapter on grammar–lexis relations.
• Empirical methods: root identification, SVM-based tokenization and tagging, transducer learning of broken plurals, and memory-based tagging.
• Applications: information retrieval and machine translation.

Kiraz calls the book essential for researchers on Arabic morphology and a sign of how far the field has come. He praises its clear explanations and references. He doubts that the DATR accounts have been shown to reach full coverage at acceptable complexity. His main complaint is that every chapter re-describes the same Arabic data. One shared description would have cut repetition and pushed authors to relate their results to each other.

Key points

• The volume runs to viii+305 pages in 15 chapters across four parts. It was edited by Soudi, van den Bosch and Neumann and priced at $159.
• Daya, Roth and Wintner learn to identify Semitic roots from surface forms. They start with Hebrew, extend to Arabic, and report precision above 80%, which they compare with average human performance.
• Diab, Hacioglu and Jurafsky apply support vector machines to tokenization, lemmatization, POS tagging and base phrase chunking. Accuracy ranges from 91.6% (chunking) to 99.1% (clitic tokenization).
• The memory-based analysis and tagging by van den Bosch, Marsi and Soudi reaches 58.1% joint accuracy. It cannot recover stems of unknown words but tags POS well.
• Larkey et al. find that light stemming is enough for IR: strip diacritics, then cut common prefixes and suffixes. Habash's bidirectional Almorgeana was tested on over one million words, both diacritized and undiacritized.

Interesting ideas

• I1. Organizing an inflection lexicon as a default-inheritance network lets shared patterns live in upper nodes, dependent forms sit lower, and syncretisms come from inference rules, so exceptions only override what differs.
• I2. Almorgeana uses one feature-set (including the stem) both to generate surface words and as the target of analysis, which gives a natural round-trip check between the two directions.
• I3. Diab et al. stretch a standard classifier beyond tagging to make segmentation decisions, recasting clitic splitting as a classification problem with very high accuracy.
• I4. Daya et al. measure their system's root recognition against average human performance on the same task, which puts a raw precision number in context.
• I5. Guessoum and Zantout build a catalogue of recurring morphological error types and score a commercial MT system against it, rather than relying on a single quality number.

For Mova

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

• I1 (tokens): The English morphology lexicon in crate en could encode regular inflection as defaults with irregular verbs and plurals as local overrides, making every produced form traceable to the node that licensed it.
• I2 (back): Word-form generation from lemma plus features should be the exact inverse of the analyzer, and an analyze→generate→compare loop over a corpus would flag inconsistencies automatically.
• I3 (tokens): Splitting English contractions and possessives (n't, 's, 'll) could be handled as an explicit per-position decision with logged features, instead of hard-coded regex exceptions.
• I4 (learn): Evaluations of hard steps such as role labeling or implicature could report a human-agreement ceiling alongside model scores, making it clear how much headroom actually remains.
• I5 (back): A typed error checklist for generated English (agreement, tense, article choice) would make the text → graph → text compressor's failures diagnosable category by category.