A Large Coverage Verb Lexicon For Arabic

Abstract

This dissertation aims at classifying verbs in Arabic using syntactic alternations for the purpose of a large coverage verb lexicon inspired by the work of Kipper Schuler (2005) on English VerbNet. The main assumption is that the diathesis alternations are meaning preserving (Levin, 1993) and can be identified cross-linguistically. The classification schema includes three steps. First, building an alternation collection for Arabic to be used as filter for classifying verbs by a. testing the availability of Levin's alternations in Arabic, b. collecting Arabic-specific alternations, c. building a set of alternations involving complement clauses, which are not concidered by Levin. Second, building coarse-grained classes (super-classes) and providing them we generalized semantic descriptions composed of semantic predicates, thematic roles and other specifications and building finer-grained classes of verbs out of them considering two main criteria: a. how verbs lexicalize the meaning components of the super-classes and b. how they realize these components in the syntactic level (as diathesis alternations). Third, encoding the resulting classes into a computer readable form to reflect the morphologic, syntactic and semantic properties by virtue of which classes are distinguished from each others. The resulting lexicon, Arabic VerbNet, adopts the general architecture and the information elements of the English VerbNet, however, it adds many features to reflect the Arabic-specific properties such as the word order which is predominantly VSO, specific phrase structures such as PPs with clausal or deverbal objects and the productive derivational behavior of verbs which forces them to be marked in some alternations. Examples of new features are new entry types such as deverbals and participles, new selectional restrictions, new phrases types, new semantic predicates and a new class type called \sibling class". Sibling classes accommodate verbs that participate in a type of alternations, that requires a morphological marking. Many of them are built automatically using the root-pattern morphology of Arabic in combination with the semantic information provided by the main classes. The evaluation of Arabic VerbNet against a concurring resource, Arabic PropBank, shows a supremacy of Arabic VerbNet with respect to the coverage (e.g. number of entries, number of frames, number of roles per class) and a high degree of matching in features like polysemy (types and number of meanings per entry), role assignment (type of roles assigned) and the frames of each class/frameset.