📄 A Large Coverage Verb Taxonomy for Arabic

Summary

Arabic lacked a broad verb lexicon that ties syntactic behaviour to meaning. Root-based and etymon-based theories group related forms but say little about argument structure, and Arabic WordNet covers only about 1400 verbs with no predicate-argument information. Mousser builds an Arabic counterpart to VerbNet. He starts from Levin's classes, adds some extended classes from Korhonen and Briscoe, and follows Schuler's VerbNet construction procedure. English class members are translated and then expanded with synonyms and hyponyms drawn from classical dictionaries. A prototype verb is fitted into every frame it allows. Other verbs either join the class or go into subclasses, depending on how well they match the prototype's thematic roles, selectional restrictions and core semantic predicates. Frames whose semantics do not fit Arabic are adjusted, and some classes are dropped, merged or newly created. The resulting resource has 173 classes, 4392 verbs and 498 frames, distributed as XML. The work shows that VerbNet's compositional descriptions can be carried into a typologically distant, morphologically rich language.

Key points

• The lexicon contains 173 classes, 4392 verbs and 498 frames. It is stored in XML, convertible into MySQL, YAML or JSON formats, and is readable in Arabic script or Buckwalter transliteration.
• Each entry is a diacritized perfective 3rd-person masculine singular form. It has child nodes for the root, one or more deverbal nouns and one or more participles, and classes can be reached through any of these.
• A preliminary study found that about 65% of English diathesis alternations also exist in Arabic. In Arabic they are often realised through morphology, while some English ones, such as the conative, are missing entirely.
• Polysemy is handled by listing a verb in several classes. For example, ǧaraā appears in run, meander, occurrence and mode of being.
• Classes linked by valency-changing derivation are connected bidirectionally, and so are their individual members. These operations are causativation, decausativation, reciprocalisation and reflexivization.
• Levin classes were adapted to Arabic:
  • debone was dropped;
  • gobble was merged into devour;
  • new classes were added, such as šahhada and saraā ('walk at night').
• The lexicon was mapped to Arabic WordNet.

Interesting ideas

• I1. Each class is defined by one prototype verb in all its frames, and candidates join the class or a subclass according to how well they match it (e.g. manner speaking split into verbs with or without topic/recipient roles).
• I2. Frame semantics are compositional predicates with event phases, so language-specific facts can be fixed by adding or removing predicates, as when smell emission becomes a change of state in Arabic.
• I3. Diathesis alternations are treated as useful evidence for class membership but not a requirement, because some alternations simply do not exist in a given language.
• I4. Derivationally related classes are linked sideways, with no parent-child hierarchy, and only when the operation is regular and applies to the whole class.
• I5. Polysemy is encoded by cross-listing a verb in several classes instead of splitting it into fine-grained senses.

For Mova

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

• I1 (world): When compiling verb classes, we could test each new verb against a prototype's role set and predicates and route mismatches to a subclass rather than forcing class membership.
• I2 (frames): VerbNet predicates could be edited per verb or per dialect at the predicate level, so that states versus changes of state feed event ordering correctly.
• I3 (learn): Alternation patterns seen in silver data could count as soft evidence for assigning verb classes, and their absence should not count as a negative.
• I4 (world): Causative/inchoative and reflexive pairs could be linked the same way in the knowledge graph, so a statement like "X broke Y" can license "Y broke" in the text's world.
• I5 (frames): Frame selection could choose among the classes a verb is listed in by checking which roles are filled in the UD tree, keeping the decision explainable.