Stemming Turkish Words with LSTM Networks

Turkish is an agglutinative language that builds words by concatenating the units called morphemes. Building words by concatenating various units together leads to sparsity problem in many natural language processing tasks such as machine translation, sentiment analysis, and information extraction because each different form of the same word is considered as a different word token. In this paper, we propose a method that can find the stems of words automatically by filtering out any derivational or inflectional suffixes attached to words. The proposed method is based on an encoder-decoder model built by recurrent neural networks. Any given word is first encoded by the neural network and then its stem is extracted by decoding it. This method has been used in problems such as tagging or machine translation so far. We obtain compatitive results compared to other Turkish stemmers. Moreover, unlike the other models, stemming could be performed without defining a rule set manually, and by just using a train set that involves word and stem pairs