Beyond the Transfer-and-Merge Wordnet Construction: plWordNet and a Comparison with WordNet

Proceedings of Recent Advances in Natural Language Processing, pages 443–452, Hissar, Bulgaria, 7-13 September 2013. Beyond the Transfer-and-Merge Wordnet Construction: plWordNet and a Comparison with WordNet Marek Maziarz & Maciej Piasecki & Ewa Rudnicka Institute of Informatics Wrocław University of Technology Poland mawroc@gmail.com maciej.piasecki@pwr.wroc.pl ewa.rudnicka78@gmail.com Stan Szpakowicz Institute of Computer Science Polish Academy of Sciences, Poland & School of Electrical Engineering and Computer Science University of Ottawa, Canada szpak@eecs.uottawa.ca Abstract Wordnets are lexico-semantic resources essential in many NLP tasks. Princeton WordNet is the most widely known, and the most influential, among them. Word- nets for languages other than English tend to adopt unquestioningly WordNet’s struc- ture and its net of lexicalised concepts. We discuss a large wordnet constructed inde- pendently of WordNet, upon a model with a small yet significant difference. A map- ping onto WordNet is under way; the large portions already linked open up a unique perspective on the comparison of similar but not fully compatible lexical resources. We also try to characterise numerically a wordnet’s aptitude for NLP applications. 1 Introduction It is hard to imagine NLP without lexico-semantic resources. The Princeton WordNet (PWN) is a powerful case in point: we have come to rely on it even in “hard-core” statistical methods of pro- cessing English texts. Wordnets for other lan- guages, which soon followed PWN,1 have usually been built by the transfer-and-merge method: the structure of PWN is copied over to the target lan- guage, the lexical material is translated, and the in- evitable differences in language typology and cul- tural background are a matter of post-processing.2 1See www.globalwordnet.org/gwa/wordnet_ table.html for an up-to-date list. 2Such differences are non-trivial even within the same language family – for example, Germanic, Romance or Slavic – and become highly significant as one moves further away from Indo-European languages. The transfer-and-merge construction allows high compatibility between PWN and the target wordnet, so also between any wordnets built in the same way. Multi-lingual NLP work benefits from dependable interlingual relations – ensured if one uses a wordnet with PWN’s structure. PWN’s se- mantic relations are undoubtedly of general utility, but they do exhibit certain “English bias”, and that – combined with the anglocentric network of con- cept underlying PWN’s synsets – is a downside of the translation method of building a new wordnet.3 The result need not be an accurate reflection of the lexico-semantic system of the target language. The translation method has another advantage: it is rather affordable, because PWN is now very complete and quite stable. To start the construc- tion of a wordnet without looking to PWN may seem a little foolhardy, but it offers certain intrigu- ing benefits. This paper looks at one of such inde- pendent projects, a wordnet for Polish. The plWordNet project aims to construct a large lexical resource (comparable in size to the largest existing wordnets, including PWN), based on few but precise principles and definitions. The goal is to achieve a faithful description of Polish while en- abling compatibility with PWN (and by …

Content truncated to 500 words