Section: New Results

Verb Classification

To help computer systems in the task of understanding and representing the full meaning of a text, verb classifications have been proposed which group together verbs with similar syntactic and semantic behaviour. For English verbs, VerbNet provides such a large scale classification but there are no similar French resource available. We investigated different ways both of automatically constructing such a resource; and of evaluating it.

Using Formal Concept Analysis (FCA), we developed a method for classifying verbs based on their (syntactic) subcategorisation information extracted from existing French lexical resources; and by translating the English Verbnet, we showed how to associate the obtained classes with semantic information represented by Verbnet's thematic role sets ([27] ). As a result, a VerbNet like classification for French verbs can be constructed fully automatically.

The FCA approach we pursued, first builds a classification based on verbs and verb features and second filters this classification using various metrics (e.g., concept probability, concept stability). We are currently comparing this approach with a clustering approach which makes use of detailed evaluation metrics [44] and uses probabilistic information to guide classification. First results are promising and outperform the state of the art methods in this domain [63] .

One important difference between the clustering and the FCA approach we experimented with is that only the second, allows a verb to belong to several classes. Since verbs are highly ambiguous, this is an important difference. To evaluate the impact of this difference on the usability of the classifications built by each of the methods, we are currently conducting a task-based, extrinsic evaluation of both classifications by analysing their impact when used in a Semantic Role Labeling task on a French corpus.