📄 Unsupervised Learning of Arabic Stemming Using a Parallel Corpus

Summary

The paper addresses how to build a stemmer for a language without hand-written rules, native experts, or annotated data. Arabic is the test case, where both prefixes and suffixes must be stripped. The authors train the Arabic stemmer from two inexpensive inputs: a parallel Arabic–English corpus of about 10K sentences and an existing English stemmer. They stem the English side and fit an adapted IBM Model 1 translation model linking Arabic stem candidates to English stems. They then re-estimate the Arabic segmentation and repeat. Each candidate is scored from its prefix, stem and suffix, so the small, frequent affix inventory carries most of the reliable signal. An optional second step adapts the stemmer to a domain using raw monolingual Arabic text. After training, no parallel text is needed. The resulting system (UNSUP) agrees with a proprietary rule- and annotation-based stemmer (GOLD) on 87.5% of cases. On TREC 2002 Arabic retrieval it raises mean average precision by 22–38% over unstemmed text and reaches 93–96% of GOLD's performance. The work shows that a cheap bilingual resource can stand in for language expertise.

Key points

• The only training resources are a ~10K-sentence UN Arabic–English parallel corpus, an English stemmer and, optionally, unannotated Arabic text. The trained stemmer runs on Arabic alone.
• Segmentation works by placing two separators inside each word, which yields prefix, stem and suffix. Very short stems of one or two letters are disallowed. The score is the product of the prefix, stem and suffix factors, and these factors can be conditioned on word length or the first and last letters.
• Three starting points were compared: no stemming, random stemming (a uniform prior) and a simple rule (strip Al at the front and p at the end). The rule-based start is a strong baseline, but the random start catches up over iterations and especially after Step 2.
• For Step 2 adaptation, a larger out-of-domain corpus (237K AFP) helped more than a smaller in-domain one (80K UN). On a cross-domain 10K-sentence AFP test set, accuracy was similar to the UN results.
• The task-based evaluation used the TREC 2002 AFP documents with 50 queries, tested with titles only, with titles plus descriptions, and restricted to 1995. UNSUP's gap to GOLD is smaller in retrieval than in direct agreement.

Interesting ideas

• I1. Scoring a segmentation through its affixes rather than its stem works well with little data: a closed, frequent set of affixes yields stable statistics even when most stems are rare or unseen.
• I2. Instead of committing to the single best candidate each round, the method keeps a probability distribution over candidates. That distribution starts uniform and gradually concentrates on one option, which avoids locking in early mistakes.
• I3. Running the same pipeline from a knowledge-free start and from a cheap rule-based start shows how much the hint actually contributes and whether learning can recover without it.
• I4. Evaluating both by agreement with a strong reference system and by a downstream task revealed that segmentations which look wrong to linguists can still serve the end task almost as well.
• I5. A parallel corpus lets analysis from a well-resourced side be transferred to a poorly resourced language, with no direct supervision in the target language.

For Mova

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

• I1 (tokens): when lemmatizing out-of-vocabulary English words, rank candidate analyses mainly by suffix statistics conditioned on word length, and give the rare stem only a smaller weight.
• I2 (learn): when several witness parsers disagree, keep weighted alternatives in the silver data and let them sharpen over retraining rounds rather than taking a hard majority vote too early.
• I3 (learn): add a "no prior knowledge" control run for self-induced repair rules, to check that the gains come from learning rather than from seeded heuristics.
• I4 (answer): measure analyzer changes not only against treebank gold but also by their effect on QA and school-test accuracy, since the two measures can diverge.
• I5 (back): when translating through the meta-world into Ukrainian, aligned sentence pairs could induce Ukrainian affix segmentation from Mova's English lemmas.