📄 Classifying Arabic Verbs Using Sibling Classes

Summary

The paper tackles a structural problem in Arabic VerbNet, a Levin-style verb lexicon modelled on the English VerbNet. In Arabic, many diathesis alternations, such as causative/inchoative or reciprocal, are marked by changing the verb's morphological pattern. The result is a distinct lexical entry. Strict top-down inheritance does not let such a form sit inside its base class or any subclass, so derived verbs end up cut off from verbs whose behaviour they largely share.

The author proposes sibling classes: independent classes that stay linked to the class they were derived from. A Java tool, the Sibling class generator (SCG), does the following:

• analyses verbs into root and pattern using regular expressions;
• picks classes with causative or change-of-state semantics;
• applies pattern rewrite rules to produce derived verbs;
• adjusts thematic roles, syntactic frames and semantic predicates to match.

After manual correction, the lexicon grew by about 44%. This shows a practical way to extend class-based lexical resources to languages with productive verb morphology.

Key points

• Starting resource: Arabic VerbNet with 202 classes, 4707 verbs and 834 frames. Each frame pairs an example sentence with its syntax and with semantic predicates that carry temporal information.
• SCG found 89 expandable classes (3005 verbs, 368 frames). Of these, 60 were causative and 29 uncausative. It generated one sibling class for each, totalling 3360 verbs, since some verbs have several causative or inchoative forms.
• Measured quality: 71% accuracy in identifying verb patterns and 82% in generating the correct derived forms.
• After manual cleanup, Arabic VerbNet reached 291 classes, 7937 verbs and 1202 frames, an expansion of about 44%.
• Many generated forms are possible but not attested in corpora. An example is causatives of animal-sound verbs. The author still treats them as valid potential lexicon entries that native speakers would understand.

Interesting ideas

• I1. Instead of forcing a morphologically derived verb into a subclass, where inheritance blocks the changed properties, give it a parallel class that keeps an explicit link to its origin; this keeps relatedness without breaking the hierarchy.
• I2. Classes are chosen for expansion by inspecting their own semantics, such as an agent role plus a CAUSE predicate, or a patient subject plus a STATE predicate, so the selection is driven by meaning rather than listed by hand.
• I3. A new frame is derived from a base frame by a fixed recipe: drop or add a role, promote the object to subject, and remove or insert CAUSE; frames unrelated to causativity, such as the instrumental, are copied over with adjusted roles.
• I4. Generation is treated as the exact inverse of analysis: the same root-and-pattern templates used to parse a verb are filled back in to produce its derived form.
• I5. The paper argues that verbs in one class tend to share derivational behaviour, so plausible but unattested forms can be predicted from class membership; it also scores pattern identification and form generation separately.

For Mova

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

• I1 (world): Our compiled verb-class layer could connect English derivational or alternation pairs (e.g. wide/widen, solid/solidify, causative vs. inchoative uses) through sibling links rather than duplicate or orphaned entries.
• I2 (frames): When mapping VerbNet roles, the same predicate-based test could flag which frames should automatically receive a derived inchoative or reflexive reading.
• I3 (frames): An explainable rule set like this could create intransitive "The clothes dried" frames from transitive ones, keeping who-did-what-to-whom consistent across alternations.
• I4 (back): Sharing one reversible rule table between morphological analysis and surface generation would keep text → graph → text round trips symmetric and easier to debug.
• I5 (learn): Unseen derived verbs could inherit a provisional class from their base. Evaluating the analysis step and the generation step separately would show where errors come from.