Unsupervised Induction of Arabic Root and Pattern Lexicons using Machine Learning

Proceedings of Recent Advances in Natural Language Processing, pages 350–356, Hissar, Bulgaria, 7-13 September 2013. Unsupervised Induction of Arabic Root and Pattern Lexicons using Machine Learning Bilal Khaliq Dept. of Informatics University of Sussex bk54@sussex.ac.uk John Carroll Dept. of Informatics University of Sussex johnca@sussex.ac.uk Abstract We describe an approach to building a morphological analyser of Arabic by inducing a lexicon of root and pattern templates from an unannotated corpus. Using maximum entropy modelling, we capture orthographic features from surface words, and cluster the words based on the similarity of their possible roots or patterns. From these clusters, we extract root and pattern lexicons, which allows us to morphologically analyse words. Further enhancements are applied, adjusting for morpheme length and structure. Final root extraction accuracy of 87.2% is achieved. In contrast to previous work on unsupervised learning of Arabic morphology, our approach is applicable to naturally-written, unvowelled Arabic text. 1 Introduction The number and diversity of human languages makes it impractical to manually craft lexicons and morphological processors for more than a very small proportion of them. Further challenges are posed by the need to deal with dialects and colloquial forms of languages. This has motivated recent increased interest in approaches to morphological analysis based on unsupervised learning. Inspired by competitions such as the Morpho Challenge, many techniques have been proposed for unsupervised morphology learning. Although these techniques are often intended to be language independent, they are often directed to a specific group of languages. Most work has aimed at sequential separation or segmentation of morphemes concatenated together in a surface word form. This type of analysis, outputting stems and appended morphemes aims to identify some kind of border between the different morphemes. However, another type of word formation consists of the interdigitation of a root morpheme with an affix or pattern template; in this case there is no boundary between morphemes, since they are rather intercalated with each other. This type of non-concatenative morphology, which is characteristic of the Semitic group of languages, has attracted far less interest for unsupervised learning. In this paper we present an approach to unsupervised learning of non-concatenative morphology, applying it to Arabic. We describe an approach to learning tri-literal roots and affix template of Arabic by first inducing root and affix lexicons. Our approach uses Maximum Entropy modelling to obtain clusters1 of words based on concatenative and non-concatenative orthographic features, and induces the lexicons from these clusters. Our data is an undiacritized version of the Quranic Arabic Corpus since we assume a realistic setting of unvowelled text, as most Arabic text is written without vowels; we chose this corpus since correct roots of each word are available, facilitating the evaluation process. The fact that the corpus contains a relatively small vocabulary of around 7000 words also simulates the scenario for most of the world’s languages of scarcity of linguistic resources and data. This paper is structured as follows: Section 2 surveys previous related work. Section 3 provides an introduction to Arabic root and pattern …

Content truncated to 500 words