MaltiMorph : a computational analysis of the Maltese broken plural

Abstract

The Maltese Broken Plurals have always been treated as a mechanism totally lacking in rules or structure. The inflection of the nouns or adjectives involved makes use of non-concatenative morphology. The traditional view is that there is no connection whatsoever between the singular form and the plural pattern in the broken plural forms. Tamara Schembri, in her B.A. thesis, asserts that the latter is not entirely irregular and attempts to discover regularities within the broken plural mechanism which might be unconscious to the native speaker in such a way that one will be able to go from the singular from to the plural form of an arbitrary noun with a reasonable degree of accuracy. Throughout this project we attempt to further her analysis of this enigmatic part of Maltese Morphology. The highly non-deterministic nature of this problem motivated us in turning to Machine Learning for a solution. We used a Connectionist Model for both the generation of the plural forms as well as the classification of the singular nouns. This entails a pattern associator network which is flanked on both sides by an encoding and decoding unit respectively. Singular nouns are fed into the network, encoded and input into the main network. The network will then output a set of features which are decoded by the decoding unit and transformed into a word which relates to the plural of the noun in question. We also created a network which can classify the singular nouns into their various categories. The main method we used to assess our results was the Cross-Validation approach. The results of the networks tend to agree with Tamara Schembri's theories. Clearly there is a certain degree of regularities, especially within certain categories. Learning was optimized to the point of convergence or very close to it; generalisation to new forms did not perform badly either when one takes into account the nature of the problem. We also think that our results can conclusively show that although we are still far off from creating a complete computational model for the Broken Plural, we have shown classification of nouns in their singular form is not impossible and can be achieved with a relatively good degree of accuracy.