A Study on Linking Wikipedia Categories to Wordnet Synsets using Text Similarity

International Conference RANLP 2009 - Borovets, Bulgaria, pages 449–454 A study on Linking Wikipedia categories to Wordnet synsets using text similarity∗ Antonio Toral∗ ´Oscar Ferr´andez† Eneko Agirre⋄ Rafael Mu˜noz† ∗Istituto di Linguistica Computazionale, Consiglio Nazionale delle Ricerche Pisa, Italy †Natural Language Processing and Information Systems Group University of Alicante, Spain ⋄Informatika Fakultatea, University of the Basque Country Donostia, Basque Country Abstract This paper studies the application of text simi- larity methods to disambiguate ambiguous links between WordNet nouns and Wikipedia cate- gories. The methods range from word overlap between glosses, random projections, WordNet- based similarity, and a full-fledged textual en- tailment system. Both unsupervised and super- vised combinations have been tried. The gold- standard with disambiguated links is publicly available. The results range from 64.7% for the first sense heuristic, 68% for an unsupervised combination, and up to 77.74% for a supervised combination. 1 Introduction Human languages are extremely rich and ambiguous resulting in the fact that the same information can be expressed with different words and linguistic struc- tures. Consequently, an ambiguous text might repre- sent several distinct meanings and a concrete mean- ing might be expressed by different ways. Therefore, language ambiguity and variability are considered as essential blocks to solve in order to overcome the bar- rier that separates the human understanding from the computer understanding. The task of detecting semantic similarity between texts addresses properly these language phenomena and also has a lot of potential applications for Natural Language Processing [15]; examples include word sense disambiguation [16], categorisation [11], summarisa- tion [4], etc. In the current research, semantic similarity is ap- plied into a methodology devoted to the automatic construction of a Named Entity Repository [18]. This method exploits the knowledge available in already ex- isting Language Resources (LR) to support procedures of lexico-semantic acquisition from Web 2.0 collabora- tive semistructured resources. Our test case for English focuses on WordNet [5] as the LR and Wikipedia as the Web 2.0 resource. The first step consists in establishing links between ∗This work has been partially funded by the EU Commis- sion (projects ICT-2007-211423 and FP6-IST-033860) and by the Spanish Government (project TIN2006-15265-C06-01). entries of both resources; the instantiable common nouns found in WordNet are mapped to Wikipedia cat- egories. Obviously, these mappings are ambiguous for polysemous nouns. Another piece of research, YAGO [17], also addresses linking WordNet to Wikipedia. However, the authors do not deal with the ambigu- ity that arises when linking both resources (ambiguous mappings are simply manually disambiguated). Our first attempt to resolve ambiguous mappings consisted in finding instances appearing both in a sense of the word in WordNet and in the mapped category in Wikipedia. This method offered perfect precision but suffered from low recall (39%) due to the small num- ber of instances present in WordNet. The alternative solution we explore in this paper consists on applying semantic similarity between the LR definitions and the mapped Wikipedia abstracts. Let us then consider our problem as a real-world testbed in which we will …

Content truncated to 500 words