Unsupervised Improvement of Morphological Analyzer for Inflectionally Rich Languages

Abstract

This paper presents an algorithm for unsupervised learning of morphological analysis and generation of inflectionally rich languages like Hindi, given a low coverage morph and a corpus of raw text. It assumes no particular theoretical model of morph, but can work with any morph that defines classes of stem that behave similarly. The morph learning algorithm uses the concept of 'observable paradigm'. The results of the algorithm are encouraging with the coverage of a primitive morph going up from 32% to about 63% and that of an advanced morph going up from 96% to about 97%.