Princeton WordNet – the prototypical wordnet (”the mother of all wordnets”) – started off as a psycholinguistic experiment on language acquisition by children. Later, it developed into a lexico-semantic database. Thus, WordNet was not originally meant to be a dictionary, but at some point began to be treated as one. It is usually presented as a network of lexicalised concepts (represented as synsets – synonym sets). In addition, many people call it and use it as a kind of ontology. Contrary to such claims, we will argue that wordnet can be modelled (and constructed) as a relational semantic dictionary in which lexical meanings function are the basic building blocks defined by a dense network of lexico-semantic relations as a primary means of their description.
In such perspective, synsets are construed as sets of lexical meanings that share lexico-semantic relations of certain types. Thus, there is no need for assigning to them a special ontological status. Relations between synsets are just notational abbreviations for beams of relations between lexical meanings. The whole construction of a wordnet is based on Minimal Commitment Principle: minimising the number of assumptions, maximising the freedom of further interpretation of wordnet structure.
In a way typical for dictionaries, all lexical properties are assigned to lexical meanings, especially non-relational elements of description such as usage examples, textual definitions or attributes like stylistic register. The properties, but also lexico-semantic relations can be based on language data in a straightforward way, e.g. by various linguistic tests verified against usage examples, not only intuitions of linguists.
In order to show the consequences of the model, we will refer to plWordNet – a wordnet of Polish – which has been consequently built on its basis. A corpus-based wordnet development process has been applied in the construction of plWordNet, i.e. large text corpora were used as a source of lexical knowledge supporting the work of lexicographers to extract, e.g., lemmas, clusters of usage examples suggesting potential meanings, multi-word expressions, distributional models revealing semantic relatedness or instances of lexico-semantic relations. The talk will be illustrated with examples and statistics zooming in on several details of the solution.