Elements of knowledge-free and unsupervised lexical acquisition

attributes that can be seen as equivalent to the semantic primitives except that these abstract attributes do not automatically receive names similar to those that can be found in manually created semantic primitive collections. In order to clarify this analogy it is feasible to make an example for the word leap. Apart from the initial attributes like ‘frog’ and ‘leg’ that will emerge from an initial analysis of a corpus where leap frequently occurs, other abstract attributes could emerge from clustering. The would result in leap becoming part of one or more clusters resembling abstract attributes. The first one, for example, could be a cluster built from the words leap, jump, run and go. This cluster could thus be seen as representing the semantic primitive move. The second cluster could comprise the words leap, big, huge, gigantic and therefore represent the semantic primitive large. A third cluster could possibly be built from the words leap, sudden, fast, jerky and represent the primitive sudden. The ‘meaning’ of the word leap could then be represented either by the semantic primitives large sudden move, which is only interesting for humans, or by the three clusters representing the attributes, which is something applications can make use of. The fuzzyness is another seemingly problematic feature of semantic primitives, because they are commonly defined to either apply (eventually positively versus negatively) or not, without fuzzyness. The meaning of the word frog therefore clearly possesses the primitive +living. However, as described in Section 1.3, this model aims at being able to analyze the structure of language, not the relation between parts of the structure and the real world. Also, whether +living is correct for the word frog depends on its usage it might have been used as a name or a concept for something. In these cases, +living might still be slightly applicable, but not as clearly as before. Further phenomena commonly encountered are linguistic relations such as antonymy, synonymy and hypernomy, which are not dependent upon the values of specific semantic categories. These abstract semantic relations are important for many NLP applications, and traditionally have been manually encoded in lexicalsemantic structures such as WordNet (Miller, 1990; Fellbaum, 1998), GermaNet (Hamp and Feldweg, 1997; Kunze and Wagner, 1999), and EuroWordNet (Vossen, 1998). One important application of these resources is to infer other relations, cf. (Richardson, 1997). In WordNet, the collection of semantic relations includes: • Hyperonymy and hyponymy also sometimes called the is-a-relation. It holds between two ‘concepts’ (which are sets of words) whenever one has a more abstract meaning than the other. • Two words are meronyms, whenever one denotes something that is part of the other word’s denotation. In WordNet, meronyms are split into several types: part-of, member-of and substance. The differences between these types depend on the type of the denotation (countable, fluid, etc.).