📄 Improved Arabic Base Phrase Chunking with a new enriched POS tag set

Summary

The paper tackles base phrase chunking (BPC) for Modern Standard Arabic. BPC is a shallow, non-recursive alternative to full parsing, which performed poorly for Arabic at the time. Diab proposes ERTS, an enriched POS tag set. It extends the 25-tag LDC reduced set (RTS) by marking definiteness, gender and number on nominals. These are features that unvowelized Arabic usually shows on the surface. Case, mood and person are left out.

Both the POS tagger and the chunker use the same SVM sequence setup (YAMCHA over TinySVM). Training chunks come from the Arabic Treebank via ChunkLink. Its English-oriented output is corrected with Arabic-specific rules.

The ERTS tagger matches RTS tagging accuracy. In chunking, ERTS gives the best overall score, 96.33 F over 10 chunk types. It beats RTS even when RTS is supplied with explicit morphological features. All conditions clearly surpass the earlier published 91.44. The broader lesson is that the design of the POS inventory strongly affects downstream syntactic processing.

Key points

• Tag set sizes: FULL ATB morphology has over 2000 tags, RTS has 25, and ERTS has 75, of which 57 actually occur in the data.
• POS tagging accuracy is 96.13% with ERTS and 96.15% with RTS. The tagger uses ±4 character n-grams from word edges, a ±2 word window, a degree-2 polynomial kernel and one-vs-all classification.
• The chunker handles ten phrase categories, which yields 21 IOB classes. The nominal and verbal categories are NP, VP and PREDP. The modifier categories are ADJP and ADVP. The function-word categories are PP, CONJP, PRTP and SBARP. INTJP covers interjections.
• Overall F scores:
  • RTS-noFeat: 95.41
  • RTS-allFeat: 95.97
  • FULL-noFeat: 96.29
  • ERTS-noFeat: 96.33, the best result
  • Adding explicit features to ERTS or FULL lowers scores slightly.
• Best per-type scores all come from ERTS conditions: ADJP 73.16, INTJP 64.29 and NP 94.92. PP exceeds 99 in every condition.

Interesting ideas

• I1. Build a middle-sized tag inventory from only those morphological features that are reliably visible in surface forms, rather than using either a coarse or an exhaustive set.
• I2. Packing features into the tag itself worked better than passing the same features as separate inputs, which seemed to add noise.
• I3. Treebank-to-chunk conversion tools carry the assumptions of their source language, so the converted output was repaired with target-specific rules (possession constructs, agreeing noun-adjective runs, clitic pronouns).
• I4. A coarser tag set slightly beat the full one, plausibly because dropping inconsistently annotated features hides gold-standard noise.
• I5. Evaluation focused separately on the phrase types the change should affect (ADJP, INTJP, NP), not only on the overall score.

For Mova

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

• I1 (tokens): The English tagger could get optional enriched tags (e.g., number or a definiteness-like determiner cue) chosen by how observable they are, then be tested on whether the parser gains.
• I2 (tree): When feeding morphology to the perceptron parser or gparse, compare a combined-tag encoding against separate feature slots instead of assuming that separate is better.
• I3 (tree): The UD dialect conversion could include a small, explainable rule layer that fixes systematic converter artifacts before training.
• I4 (learn): When silver or gold labels disagree on some attribute, try masking that attribute during training and check whether the error rate falls.
• I5 (learn): Each change could report metrics on the constructions it targets, alongside unaffected ones as a negative control.