Unsupervised Learning for Persian WordNet Construction

Proceedings of Recent Advances in Natural Language Processing, pages 302–308, Hissar, Bulgaria, 12-14 September 2011. Unsupervised Learning for Persian WordNet Construction Mortaza Montazery NLP Lab, School of ECE, College of Engineering, University of Tehran, Tehran, Iran Mortaza.gh@gmail.com Heshaam Faili NLP Lab, School of ECE, College of Engineering, University of Tehran, Tehran, Iran hfaili@ut.ac.ir Abstract In this paper we introduce an unsupervised learning approach for WordNet construction. The whole construction method is an Expecta- tion Maximization (EM) approach which uses Princeton WordNet 3.0 (PWN) and a corpus as the data source for unsupervised learning. The proposed method can be used to construct WordNet in any language. Links between PWN synsets and target language words are extracted using a bilingual dictionary. For each of these links a parameter is defined that shows probability of selecting PWN synset for target language word in corpus. Model para- meters are adjusted in an iterative fashion. In our experiments on Persian language, by se- lecting 10% of highly probable links trained by the EM method, a Persian WordNet was obtained that covered 7,109 out of 11,076 dis- tinct words and 9,427 distinct PWN synsets with a precision of more than 86%. 1 Introduction One of the most important challenges with re- spect to Natural Language Processing is the exis- tence of ambiguity in different levels of natural language. Word sense ambiguity is one of these ambiguities. One solution for dealing with these problems is to generate knowledge repositories where human knowledge about natural language can be encoded. WordNet is a rich repository of knowledge about words that has been con- structed to deal with word sense ambiguity prob- lem . The first WordNet was constructed for English language in Princeton University under direction of George A. Miller (Fellbaum, 1998). English words in four categories noun, verb, adjective and adverb have been grouped into sets of cogni- tive synonyms that are called synset. By proving of usefulness of Princeton WordNet (PWN), construction of WordNet for other languages has been considered. Two great efforts in construct- ing WordNet for other languages are Euro- WordNet (Vossen, 1999) and BalkaNet (Tufis¸, Cristea, & Stamou, 2004). The former deals with European's languages such as English, Dutch, German, French, Spanish, Italian, Czech and Es- tonian. The second one deals with languages from Balkan zone such as Romanian, Bulgarian, Turkish, Slovenian, Greek and Serbian. Manual construction of WordNet is a time consuming task and requires linguistic know- ledge. The estimation of the average time for building a lexical entry depends on the polysemy of the words in the synsets, on the available lexi- cal resources and definitely on the WordNet building tools. Thus automated approaches for WordNet construction or enrichment have been proposed to facilitate faster, cheaper and easier development. In this way several automatic me- thods have been proposed for constructing WordNet for Asian languages such as Japanese, Arabic, Thai and Persian that use PWN and other existing lexical resources. In (Shamsfard M. , 2008) a semi-automated method has been proposed for …

Content truncated to 500 words