📄 Revisiting Arabic Semantic Role Labeling using SVM Kernel Methods

Summary

This demonstration paper describes a planned, not yet finished, system that labels semantic roles in Modern Standard Arabic text. It follows the earlier work of Diab, Moschitti and Pighin. The authors argue that Arabic tools are usually ported from English, and that Arabic has its own traits: rich inflection carried by diacritics that are often omitted, flexible word order, complex idafa noun phrases, and frequent pro-drop. The pipeline splits SRL into two steps: finding argument boundaries and classifying the arguments.

Every pairing of a predicate with a parse-tree node becomes a positive or negative training example for SVM-light-TK, used with a polynomial kernel. The first version uses the "standard" English-derived features, and Arabic-specific features are planned for later. Training data is a small set of news-blog sentences labeled for ARG0 and ARG1.

No results are reported yet. The authors expect to match the published scores of Diab et al. and ask the community for design feedback. The paper matters mainly as a clear outline of a reproducible feature-based SRL setup for a morphologically rich language.

Key points

• The corpus has just over 100 sentences from news-oriented Arabic blogs. They were processed with the AMIRA 2.0 POS tagger and BP chunker, then annotated in PropBank-like style:

  • only ARG0 and ARG1 are labeled for now;
  • ARGM phrases are marked for future use.

• All corpus sentences have explicit subjects, and nearly all are SVO. Yet the Arabic Treebank has SVO and VSO at 35% each, and about 30% of its sentences are pro-dropped.

• The learner is SVM-light-TK, mostly run as a plain SVM-light with a polynomial kernel. The tree-kernel capacity is kept for later extension.

• The features are the standard set:

  • phrase type, predicate lemma, head word and its POS;
  • position and voice;
  • full, partial and no-direction path;
  • first and last word/POS of the constituent;
  • subcategorization and syntactic frame.

• The reference targets come from Diab et al. (2007), scored with the CoNLL evaluator, and from Diab et al. (2008):

  Source	Task	F1 (dev / test)
  Diab et al. 2007	Argument classification	77.85 / 81.43
  Diab et al. 2007	Boundary detection	93.68 / 94.06
  Diab et al. 2008 (with Arabic features)	Combined	82.17

Interesting ideas

• I1. Each candidate (predicate, tree node) pair is labeled positive only if the node's subtree covers exactly one argument's words. The same example pool, regrouped, then trains the role classifier as well.
• I2. Three path variants are used: the full directed tree path, the partial path up to the common ancestor, and the path with directions stripped. Together they give coarse-to-fine views of how an argument attaches to its predicate.
• I3. Features are added in stages: first a language-neutral standard set, then language-specific morphological ones (number, gender, definiteness, diacritized lemma). This isolates how much the specialized features contribute.
• I4. Role inventories are extended by running several one-vs-all classifiers, one per argument type, and choosing the highest-scoring label.
• I5. The authors openly note that their tiny corpus lacks pro-drop and word-order variety, so the evaluation would not test those phenomena.

For Mova

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

• I1 (learn): Silver SRL data could be turned into exact-span positive and negative pairs over UD subtrees, giving one reusable pool for both a span detector and a role classifier.
• I2 (frames): These variants over UD dependency paths could serve as explainable evidence when mapping tree positions to VerbNet roles, with each decision traceable to a concrete path.
• I3 (learn): The same staging could be applied to role assignment, measuring what morphology adds over pure tree features as a built-in ablation control.
• I4 (frames): Per-role scorers keep each role decision inspectable and let new roles be added without retraining the others.
• I5 (answer): Test sets for readers and students could be audited for phenomenon coverage, such as implicit subjects or non-canonical order, before scores are trusted.