Development of the Hungarian WordNet Ontology and its Application to Information Extraction

This paper presents an outline of the construction process of the Hungarian WordNet Ontology, and the description of an information extraction application utilizing the ontology. and MorphoLogic) in a 3-year project funded by the European Union ECOP program (GVOP-AKF-2004-3.1.1.) The Princeton WordNet (WN) linguistic ontology ([1]) has become a standard and an invaluable semantic resource within the natural language technology community. The WN lexical semantic network consist of nodes called synsets (sets of synonymous words that are interchangeable in a context) corresponding to linguistic concepts and connecting edges corresponding to semantic relationships like hypernymy (is-a relationship), meronymy (part-of relationship), antonymy etc. The EuroWordNet (EWN) project ([5]) extended the WN architecture to a multilingual level, with the synsets of the English WN serving as interlingua (ILI) among the concepts of the various other languages. A common starting set (Common Base Concepts) was implemented in each participating language and then was expanded individually in a top-down manner by each partner. In addition to the 11 European languages covered by EWN, the BalkaNet project ([4]) several years later introduced connected Wordnets for 5 more Southeast-European languages. The Hungarian WordNet (HuWN) project follows the BalkaNet project's resources: Princeton WordNet 2.0 as ILI, 8500 base concept synsets as a starting point and the VisDic XML-based ontology/dictionary editor. We also decided to integrate existing semantic resources into HuWN: on the one hand, we tried to map each Hungarian synset to a sense in the EKSz Hungarian explanatory dictionary, in order to obtain definitions, and on the other hand, for each verbal synset we registered corresponding entries in our existing verb frame description lexicon. For nouns, adjectives and adverbs, the work followed the so-called expansion approach, which means we took English WordNet as a starting point for our concept networks. We used several machine-translation heuristics ([2]) to obtain a rough translation of the English synsets, which were then all manually examined, corrected and extended as necessary, with adaptation to the semantic conditions of the Hungarian language. For verbs, this approach proved to be unsustainable because of the major differences between the English and Hungarian verb typing systems. In this case, we used a mixed approach, by translating only a subset of the common concepts and creating the rest from scratch from frequent items in corpora. We also added new semantic relations and the so-called nucleus structure ([3]) in order to represent aspects of verb meanings unique to …