A Semantic Relatedness Measure Based on Combined Encyclopedic, Ontological and Collocational Knowledge

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 1397–1402, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP A Semantic Relatedness Measure Based on Combined Encyclopedic, Ontological and Collocational Knowledge Yannis Haralambous Institut Télécom – Télécom Bretagne Département Informatique UMR CNRS 3192 Lab-STICC Technopôle Brest Iroise CS 83818, 29238 Brest Cedex 3, France yannis.haralambous@telecom-bretagne.eu Vitaly Klyuev University of Aizu Aizu-Wakamatsu Fukushima-ken 965-8580, Japan vkluev@u-aizu.ac.jp Abstract We describe a new semantic relatedness mea- sure combining the Wikipedia-based Explicit Semantic Analysis measure, the WordNet path measure and the mixed collocation index. Our measure achieves the currently highest results on the WS-353 test: a Spearman ρ coeffi- cient of 0.79 (vs. 0.75 in (Gabrilovich and Markovitch, 2007)) when applying the mea- sure directly, and a value of 0.87 (vs. 0.78 in (Agirre et al., 2009)) when using the predic- tion of a polynomial SVM classifier trained on our measure. In the appendix we discuss the adaptation of ESA to 2011 Wikipedia data, as well as vari- ous unsuccessful attempts to enhance ESA by filtering at word, sentence, and section level. 1 Introduction 1.1 Semantic Relatedness and Corpora Semantic relatedness describes the degree to which concepts are associated via any kind of semantic rela- tionship (Scriver, 2006). Its evaluation is a fundamen- tal NLP problem, with applications in word-sense dis- ambiguation, text classification, information retrieval, automatic summarization and many other fields. In re- cent decades, a great variety of relatedness measures have been defined, based on corpora such as Wikipedia, Wiktionary, WordNet, etc. Wikipedia is one of the most successful collaborative projects of all time. By a constantly growing number of additions, corrections and verifications, its contents grows in both quantity and quality, and is considered by many linguists as the corpus they had always dreamed of (Medelyan et al., 2008). By measuring the normalized tfidf values of words in a page, we can consider the page to be a weighted vector in the space of words. Inverting the matrix of these vectors we obtain weighted vectors of words in the space of pages. As every page deals with a single topic, we consider these vectors as being concept vec- tors. The ESA (Explicit Semantic Analysis) measure between two words is obtained by taking the cosine of their concept vectors (Gabrilovich and Markovitch, 2007). Unlike Wikipedia, WordNet (Miller, 1995), a semi- formal lexical ontology (Huang et al., 2010), has a fine and carefully-crafted ontological structure: word senses are represented by sets of synonyms (“synsets”), and there is a graph structure on synsets based on hy- pernymic relations. Several WordNet-based semantic relatedness measures have been defined, based on dis- tances in the hypernymic graph, and often combined with word distribution in sense-tagged corpora. 1.2 Evaluation of Results, WS-353 Test (Finkelstein et al., 2001) introduce WS-353, a semantic relatedness test set consisting of 353 word pairs1 and a gold standard defined as the mean value of evaluations by up to 17 human judges. Although this test suite con- tains some quite …

Content truncated to 500 words