📄 Lexicon Acquisition for Dialectal Arabic Using Transductive Learning

Summary

The paper tackles building a part-of-speech lexicon for Levantine Colloquial Arabic, a dialect with almost no annotated resources. For an unsupervised trigram HMM tagger trained with EM, the lexicon decides which tags each word may take, and a loose lexicon badly hurts accuracy. The only inputs are raw transcripts of Levantine speech and a morphological analyzer built for MSA. Run over the dialect wordlist, that analyzer leaves about a quarter of the words without any analysis.

The authors treat filling in these missing POS-sets as transductive learning: the unlabeled words are known in advance, so the goal is to label exactly that fixed set well. Each word's set of tags is predicted as one combined class. They compare Transductive SVM, Spectral Graph Transducer, a new Transductive Clustering method and an inductive SVM.

Every learned lexicon gives a better tagger than simple heuristic baselines. Error analysis shows that mistakes inherited from the analyzer hurt EM training more than they hurt lexicon learning. Expert lexicons and supervised training serve as reference points for how much human annotation would pay off.

Key points

• Only two resources are assumed: raw Levantine transcriptions and an MSA morphological analyzer. 23% of Levantine word types get no analysis from it.
• The tagset has K=20 tags. The compound-label framing (each distinct POS-set is one class, one-vs-rest, argmax) worked better than one binary classifier per tag.
• The algorithms compared are TSVM, SGT, Transductive Clustering (TC) and an inductive SVM (ISVM). Baselines either assign all open-class tags or all tags to unanalyzable words.
• The learned lexicons have average POS-set sizes of about 1.8–1.9 tags, against 3.39 (open class) and 5.17 (all tags) for the baselines. Precision is higher, at the cost of some recall.
• Correcting analyzer errors with oracle POS-sets lifted TSVM tagging accuracy to 66.54%. Reference points:
  • an expert lexicon of 2450 items with EM training reaches 74.45%;
  • supervised training on about 2.2k sentences (16k tokens) reaches 82.93%;
  • learning curves show how accuracy grows with annotation size.

Interesting ideas

• I1. When the unlabeled items form a fixed, known list, optimize the labeling of that list directly (transduction) instead of learning a general classifier; the full test set then helps shape the decision.
• I2. Predicting a whole tag set as one compound label captures which tags tend to co-occur, but it cannot propose combinations never seen in training; this trade-off is worth weighing for any set-valued output.
• I3. A lexicon's allowed tags act as hard zeros that EM can never undo, so a tighter, precise lexicon matters more than a high-recall one.
• I4. Swapping oracle data into one pipeline stage at a time (lexicon learning vs. EM training) shows where upstream errors hurt most.
• I5. Learning curves comparing "annotate N lexicon entries" against "annotate N sentences" make annotation cost explicit and show when human labeling becomes worth it.

For Mova

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

• I1 (tokens): Filling in missing lemma or POS entries for the closed vocabulary of a given corpus could use transductive labeling over the whole word list rather than a generic out-of-vocabulary model.
• I2 (world): Assigning sets of WordNet/VerbNet classes to a verb could model common class bundles as atomic labels, with a per-class fallback for rare combinations.
• I3 (tokens): Keep the tagger's allowed-tag dictionary lean and precision-oriented, since adding rare but correct tags makes downstream disambiguation harder.
• I4 (learn): The same per-stage oracle substitution can attribute errors between tagger, parser and frame extraction, so repair effort goes where it gives the biggest gain.
• I5 (learn): When deciding between LLM-teacher tiers and silver data, plot accuracy against the size of each annotation type to choose the cheapest effective supervision.