📄 Word Segmentation of Informal Arabic with Domain Adaptation

Summary

Splitting clitics off Arabic words helps downstream tasks such as parsing and translation. However, existing segmenters either target formal Modern Standard Arabic (MSA) only or depend on hand-built, dialect-specific morphological analyzers. The authors extend the character-level CRF segmenter of Green and DeNero (2012) in three ways:

• Orthographic rewrites: an extra label marks characters whose spelling changes once a clitic is split off.
• New features: these were chosen by studying errors on Egyptian development data.
• Domain adaptation: newswire, broadcast news and Egyptian forum text are combined using Daumé's feature space augmentation.

Training and evaluation use the newswire ATB, the Broadcast News treebank (BN) and the BOLT Egyptian Arabic Treebank (ARZ). On Egyptian test data the system reaches 92.09% F1, against 91.60% for the dialect-specific MADA-ARZ. It is at least as good as MADA and MADA-ARZ almost everywhere and decodes at least seven times faster. Because it needs no analyzer or lexicon, any dialect with segmented training data can be added cheaply. The code ships in the Stanford Word Segmenter.

Key points

• The model is a third-order linear-chain CRF over characters with labels {I, O, B, F, Rew}. Rew covers three rewrite cases: restoring the article al- after li-, turning t back into the word-final h, and swapping y/ā.
• The added features are word length and the character's position inside the word. The first and last two characters of a word are also tied jointly to the boundary labels, which targets the Egyptian negation circumfix mā-…-sh.
• Domain adaptation copies every feature into one shared version plus one version per domain (N = 3). On the development set, this alone raised ARZ F1 from 88.57 to 90.37; the new features brought it to 92.10.
• On test data the system wins on F1 in all three domains. MADA-ARZ remains ahead on TEDEval for ARZ (93.33 vs 92.32), because it applies more spelling normalizations.
• The authors sampled 100 ARZ errors and sorted them:
  • 33 come from gold-data problems (7 typos, 26 annotation inconsistencies).
  • 36 come from features that are too local.
  • 31 would need wider context, for example wlā, -nā and -y ambiguities.

Interesting ideas

• I1. A single shared feature plus a per-domain duplicate lets one model learn both general and genre-specific weights with almost no engineering, and earlier domains do not degrade.
• I2. Spelling changes are folded into the segmentation label set, so the tagger both splits a token and states how each piece should be rewritten in one decoding pass.
• I3. Features that link a word's beginning and end catch dependencies a sliding character window misses. Examples are circumfixes, or an affix that only makes sense given the rest of the token.
• I4. Scoring is done on F1 over segment boundaries rather than accuracy, because over 80% of tokens are trivially single-segment. A fixed sample of errors is also sorted into gold noise, missing features and missing context.
• I5. To compare segmentations whose spelling differs from gold, each sentence is turned into a flat tree and scored with an edit-distance tree metric (TEDEval).

For Mova

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

• I1 (tree): the perceptron parser could be trained on mixed treebanks (news, fables, school texts) with domain-tagged feature copies instead of separate models per genre.
• I2 (tokens): the English tokenizer could emit rewrite labels for contractions and colloquial forms ("won't" → will + not, "gonna" → going + to), keeping each rewrite explainable.
• I3 (tokens): morphology and lemma decisions could use joint prefix/suffix features, such as un-…-able, to rule out locally plausible but globally impossible splits.
• I4 (learn): evaluations should report scores on the rare decisions, and periodic error triage should decide whether to fix silver data, add features, or pass context down the pipeline.
• I5 (tree): when tokenizer output and treebank tokens disagree, TEDEval-style scoring can still compare parses fairly instead of discarding mismatched sentences.