From efficiency to portability: acquisition of semantic relations by semi-supervised machine learning

Numeric approaches to the corpus-based acquisition of lexical semantic relations offer robust and portable techniques, but poor explanations of their results. On the other hand, symbolic machine learning approaches can infer patterns of a target relation from examples of elements that verify this relation; the produced patterns are efficient and expressive, but such techniques are often supervised, i.e. require to be (manually) fed by examples. This paper presents two original algorithms to combine one technique from each of these approaches, and keep advantages of both (meaningful patterns, efficient extraction, portability). Moreover the extraction results of these two semi-supervised hybrid systems, when applied in an illustrative purpose to the acquisition of semantic noun-verb relations defined in the Generative Lexicon framework (Pustejovsky, 1995), rival those of supervised methods.