Hungarian is one of the agglutinative languages which have a very rich, complex morphology. Words are usually ambiguous, and they can have various meanings and possible structures which are described by morphological analyses. The task of a morphological disambiguator is to determine the correct analysis of words based on their context.
Deep learning was successfully used in several natural language processing tasks and achieved state-of-the-art results, but it is not investigated much in morphological disambiguation.
In this thesis, I design and evaluate a Hungarian morphological disambiguator applying deep learning methods. I develop and train both recurrent and convolutional neural networks to accomplish the task. I examine Hungarian morphological analyses and create an appropriate method for word analysis representation considering the characteristics of Hungarian analyses and various special cases.
During the evaluation, the developed systems achieved 97% accuracy in morphological disambiguation of words and disambiguated 73% of whole sentences correctly.