String Transformation Based Morphology Learning

There are several morphological methods that can solve the morphological rule induction problem. For different languages this task represents different difficulty levels. In this paper we propose a novel method that can learn prefix, infix and suffix transformations as well. The test language is Hungarian, and we chose a previously generated word pair set of accusative case for evaluating the method, comparing its training time, memory requirements, average inflection time and correctness ratio with some of the most popular models like dictionaries, finite state transducers, the tree of aligned suffix rules and a lattice based method. We also provide multiple training and searching strategies, introducing parallelism and the concept of prefix trees to optimize the number of rules that need to be processed for each input word. This newly created novel method can be applied not only for morphology, but also for any problems in the field of bioinformatics and data mining that can benefit from string transformations learning.