Unsupervised morphological analysis using tries

Morphological analysis or decomposition studies the structure, formation, functionof words, identifies the morphemes (smallest meaning-bearing elements) of thelanguage and attempts to formulate rules that model the language. It is widely used indifferent areas such as speech recognition, machine translation, information retrieval,text understanding, and statistical language modeling. Considering that the naturallanguage processing applications are dealing with large amounts of data, it is not feasibleto use linguists to analyze text corpus by hand, the complexity and real timeprocessing requirements leads to automated morphological analysis. As an alternativeto the hand-made systems, there exist algorithms that work unsupervised manner andautonomously do morphological analysis for the words in an unannotated text corpus.In this thesis, an unsupervised learning algorithm is proposed to extract informationabout the text corpus and the model of the language. The proposed algorithmconstructs a trie that consists of characters and the occurrences of the words as nodes.The algorithm then detects roots of the given words by examining the occurrences inthe path of the word. When the root is revealed, the algorithm creates a new trie fromthe affix parts, left after the root for each word. The algorithm continues recursivelyuntil there is no affix left to process. Experimental results on three languages (Finnish,English and Turkish) show that our novel algorithm performs better than most of theprevious algorithms in the field and gives promising results.