Orð og tunga - 01.06.2012, Blaðsíða 19

Mattheio Whelptcm: From human-oriented dictionaries
9
tained 50,000 synsets. Pedersen et al. (2009) report on the problems of
converting a human-use dictionary to a lexical semantic database and
the limitations of using the classical lexical semantic relations of the
Princeton WordNet. A good example of both problems comes in the
discussion of the word butik ,shop':
For example all hyponyms of butik (shop) inherit the involved agent
handlende (shopkeeper). Thus, the DanNet editor is prompted to
identify the involved agent of the more restricted hyponym: that
the shopkeeper of a pharmacy is a pharmacist, the shopkeeper of a
bakery is a baker and so on. Such information is only rarely speci-
fied in DDO definitions (although sometimes provided implicitly as
examples of word formation), but this information is seen as highly
relevant in a wordnet.
(Pedersen et al. 2009: 273)
The DanNet developers address two issues in this passage.
The first issue is the gap between a traditional dictionary and an
nlp resource: the traditional dictionary entries of ddo leave implic-
it the relation between a pharmacy and a pharmacist, a bakery and
a baker, and so on, because the inference of such a relation can be
left to the world knowledge and common sense of a human user; a
computer however will not automatically make such inferences. The
DanNet developers therefore state the shopkeeper for every subkind
of shop, where such an entity is lexicalised. The manual process of
adding such information is streamlined by exploiting the inheritance
relation: an algorithm is set to prompt the developer for "missing in-
formation" that can be inferred on the basis of established relations;
so, if a relation is explicitly stated for a hypernym, then it is likely
that all hyponyms will have a specific equivalent of this relation: if
all shops have a shopkeeper (involved agent) and a bakery is shop
(hyponymy), then a bakery will have a shopkeeper, for whom the
language may well have a specialised lexical item.
The second issue implied in the quoted passage concerns the lim-
its of the original WordNet relation set. Notice that the information
being added to DanNet here (involved agent) is not a relation in the
original WordNet: speakers of a language implicitly understand a
systematic relation between an entity and its typical owner or user;
this information is therefore added systematically to DanNet. In fact,
the developers of DanNet extend tlie classical relations of WordNet,
following the work of the computational semanticist, Pustejovsky,