A Semi-Supervised Method for Arabic Word Sense Disambiguation Using a Weighted Directed Graph

International Joint Conference on Natural Language Processing, pages 1027–1031, Nagoya, Japan, 14-18 October 2013. A Semi-Supervised Method for Arabic Word Sense Disambiguation Using a Weighted Directed Graph Laroussi Merhbene LATICE / Faculty of sciences of Monastir, Monastir, Tunisia Arous- si_merhben@hotmail.co m Anis Zouaghi LATICE, ISSAT Sousse, University of Sousse, Tu- nisia Anis.zouaghi@gmail .com Mounir Zrigui LATICE, Faculty of sciences of Monastir, Monastir, Tuni- sia mounir.zrigui @fsm.rnu.tn Abstract In this paper, we propose a new semi- supervised approach for Arabic word sense disambiguation. Using the corpus and Arabic Wordnet1, we define a method to cluster the sentences containing ambiguous words. For each sense, we generate a cluster that we use to construct a semantic tree. Furthermore, we construct a weighted directed graph by match- ing the tree of the original sentence with se- mantic trees of each sense candidate. To find the correct sense, we use a similarity score based on three collocation measures that will be classified using a novel voting procedure. The proposed method gives a high rate of re- call and precision. 1 Introduction The human language is so complex to be learned. The syntactic form of words, the relation be- tween a specific word form and its meaning are the basic parts of an intelligent system for the natural language processing. In this work we aim to solve the task of identi- fying the sense of the ambiguous word. This task is called Word Sense Disambiguation (WSD), which is one of the oldest problems in natural language processing (NLP) (Agirre and Edmond, 2006). This work is part of a general frame-work of Arabic speech (Zouaghi, 2008). In this work, we combine a supervised and an unsupervised method for Arabic word sense dis- ambiguation. The innovative part in this work is 1 Arabic Wordnet is a concept dictionary with map- pings between word definitions. the construction of a semantic tree for each sense of the ambiguous word. Also we define a voting procedure that gives a weight for the score meas- ure. This paper is organized as follows. Section two describes the proposed method. The experi- mental results are described in section three. Fi- nally this paper is concluded in section four. 2 Proposed method We propose a semi-supervised method for Ara- bic word sense disambiguation. For the unsupervised part of the proposed method, we use Arabic Wordnet (Black et al., 2006) and the corpus to construct sense clusters (group of sentences) characterizing a specific sense of the ambiguous word. Furthermore we construct a semantic tree for each sense of the ambiguous word. The disambiguation procedure is based on the step of matching the semantic tree with the tree of the original sentence. We use a score measure (based on three collocation measures) to find the closest semantic tree to the tree of the sentence to be disambiguated. The supervised part uses a voting procedure that will rank collocation measures during a clas- sification task. The sense given by the measure having the highest rank will be attributed …

Content truncated to 500 words