📄 Robust Part-of-speech Tagging of Arabic Text

Summary

The paper tackles part-of-speech tagging for Modern Standard Arabic. Its rich clitic morphology and the many words that never appear in training make simple taggers brittle. The authors start from MD-AMIRA, a fast cascaded pipeline inside the MADAMIRA suite. It runs three linear SVM steps in sequence: character-level clitic segmentation, a classifier that restores underlying forms, and token-level tagging with a fine-grained tagset called ERTS2. To this baseline they add external knowledge, and call the result MD-AMIRA+EX. The additions are candidate tags and morphological attributes from the ALMOR analyzer, optional soft output constraints, and word-cluster IDs (Brown and word2vec K-means) learned from a large unlabeled newswire corpus. On the held-out ATB newswire test set the enhanced tagger reaches 97.15%, beating MD-MADA's 96.91% while running about 30% faster. Accuracy on unseen words also rises sharply. On cross-genre text it trails MD-MADA slightly, mainly because of segmentation mistakes. The work shows that a simple linear model, fed well-chosen resource-derived features, can rival a heavier analysis-and-ranking system.

Key points

• Results on the ATB1-2-3 held-out test set:

  Model	Overall accuracy	Unseen-word accuracy	Speed (tokens/sec)
  MD-AMIRA	96.78%	73.38%	~2415
  MD-AMIRA+EX	97.15%	89.22%	~1395
  MD-MADA	96.91%	85.28%	~1050

  With broad tags, MD-AMIRA+EX reaches 98.24%.

• Results on cross-genre data: MD-AMIRA+EX scores 94.38% versus 94.71% for MD-MADA. With gold tokenization MD-AMIRA+EX rises to 95.4%, and on the main test set gold tokenization gives 97.64%.

• Segmentation is a character-level IOB task with prefix, word and suffix segments. Normalization is a separate token-level classifier with three labels: CIP (trailing t→p), CIY (trailing y/A→Y) and NA.

• The biggest single gain came from using a few top candidate tags from ALMOR as features. ALMOR's morphological attributes and the soft constraints added little on top of those tags. Cluster IDs still helped, especially for unseen words and other genres.

• On proper nouns MD-AMIRA+EX makes fewer errors than MD-MADA:

  • mistagged as adjective: 69 vs 125;
  • mistagged as verb: 50 vs 107.

Interesting ideas

• I1. Instead of letting a morphological analyzer hard-filter the output, the paper feeds its top few candidate tags to the classifier as ordinary features, which gave most of the benefit; once that was done, extra constraints became redundant.
• I2. Splitting clitics is treated as per-character B/I/O labelling, followed by a separate small classifier that restores the underlying form of a stem changed by affixation.
• I3. Cluster IDs learned from plain unlabeled text give a linear model a cheap handle on words it never saw in labelled data, and this help persisted across genres.
• I4. The evaluation re-runs the tagger on gold tokenization, reports unseen-word accuracy separately and adds a cross-genre test set, which pins down how much error is inherited from the upstream step.
• I5. A very fine-grained tagset is predicted once and then collapsed deterministically into broad classes, so a single model serves several consumers.

For Mova

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

• I1 (tokens): our English tagger could take lexicon or morphology candidate tags as features rather than hard masks, keeping decisions soft and the evidence traceable.
• I2 (tokens): the same two-step design fits English contractions and possessives ("won't" → will + n't, "'s"), keeping segmentation and form restoration as separately explainable steps.
• I3 (tokens): Brown-style cluster features could back off the perceptron tagger on rare or out-of-domain words without making it a black box.
• I4 (learn): every pipeline stage could be scored both on gold and on predicted input, so regressions are attributed to the right component.
• I5 (tree): predicting rich XPOS-style tags and mapping them down to UPOS lets the parser and treebank conversion share one tagger output.