📄 Automatic Tagging of Arabic Text: From Raw Text to Base Phrase Chunks

Summary

The paper addresses the lack of fully automatic basic processing tools for Arabic. At the time, clitic segmentation and POS annotation depended on annotators manually choosing among candidate analyses from Buckwalter's AraMorph. The authors build a three-stage pipeline on the Arabic TreeBank (AFP newswire, non-vocalized, Buckwalter transliteration):

• Tokenization: splitting proclitics and enclitics off stems.
• POS tagging: using the 24-tag collapsed treebank tagset.
• Base phrase chunking: finding non-recursive phrases.

All three stages are framed as sequence tagging and solved with the same multi-class SVM setup adapted from English work. Only the input units, the context window and the features change between tasks, and no feature is specific to Arabic. The results are:

• tokenization F-score of 99.12
• POS accuracy of 95.49%
• chunking F-score of 92.08

All three clearly beat simple baselines. When the identical method was trained on English data of similar size, it scored about the same, which supports the claim that the method does not depend on the language. The authors present these as the first reported automatic results for these three Arabic tasks, a practical starting point for parsing and other downstream work.

Key points

• Data: the Arabic TreeBank (Arabic Penn TreeBank 1 v2.0) contains 734 AFP articles, about 140k words or 168k tokens. It was split into 4000 training, 119 development and 400 test sentences out of 4519. The setup used an SVM with a degree-2 polynomial kernel and C=1, and scoring used the CoNLL evaluation tools.
• SVM-TOK: the tokenizer labels each character with one of six segment tags (B-PRE1, B-PRE2, B-WRD, I-WRD, B-SUFF, I-SUFF) using a ±5 character window. It reaches F 99.12, against 88.62 for a rule baseline (RULE) and 93.71 for rules plus a 47,261-entry dictionary from Buckwalter's dictStem (RULE+DICT). The determiner Al is not split off, following treebank practice.
• SVM-POS: the tagger uses a ±2 token window with character n-grams and a token type feature (alphabetic or numeric). It achieves 95.49% accuracy versus 92.2% for a most-frequent-tag baseline. About half of its errors are NN/JJ confusions, partly caused by inconsistent gold annotation.
• SVM-BP: the chunker covers 9 phrase types with IOB labels (19 classes), using training data produced by flattening treebank trees with ChunkLink. It scores F 92.08 overall, with VP at 97.6 and PP at 98.4.
• English controls: the same tagger and chunker trained on English data of matched size reached 94.97% POS accuracy and chunking F 93.05.

Interesting ideas

• I1. Segmentation as character-level BIO tagging. Morphological segmentation becomes an ordinary sequence-labeling task with no handcrafted analyzer. This could transfer to other clitic-heavy or agglutinative languages, including splitting Ukrainian particles or handling compounds.
• I2. One learner, three tasks. A single classifier setup serves every level, with only the inputs, window and features changing. This makes the pipeline cheap to build and maintain, and easy to port to a new language once a treebank exists.
• I3. Matched-size cross-language control. The authors retrained on English with the same amount of data to judge whether the Arabic scores are good. This gives a fairer reference point than comparing against English systems trained on much larger corpora.
• I4. Deriving chunk data from existing trees. Flattening treebank trees into chunk labels (here with ChunkLink) yields chunking training data without any new annotation effort.
• I5. Error analysis that checks the gold data. Tracing the NN/JJ errors back to inconsistent gold labels shows that part of the remaining error lies in the annotation, not the model. This kind of check is worth doing before spending effort on model tuning.