Using deep learning algorithms character-based word generation

The aim of this study is to develop a character-based language model using deep learning algorithms and to produce meaningful words in accordance with Turkish grammar rules. Text generation is used to create text in natural language by imitating the way computers communicate. In natural language, text data exists side by side and in the form of a holistic arrangement of related words or phrases arranged in chronological order. The order and meaning of the characters and words in the text are of great importance. For this reason, deep learning algorithms that learn by remembering past information should be used in order to learn the relationships between characters or words in words, sentences and text fragments to be created. Recursive artificial neural networks, one of the deep learning algorithms, are used effectively in this field because they give successful results in creating sequential patterns by remembering past information. In the study, meaningful words in accordance with Turkish grammar rules are produced by using LSTM, GRU, Encoder-Decoder and attention architecture. In the study, especially the Encoder-Decoder and attention model are emphasized. The reason is that other models (LSTM/GRU) fail to remember the relationships between characters as the length of the input and output data increases. In order to overcome these problems, the Encoder-Decoder and Attention architecture is used, which gives more successful results in longer input and output strings. In the study, each model is operated at different threshold values of the temperature sampling method at 100, 300, 500 epoch values. LSTM model 100 epoch and temperature sampling method best result at 0.5 threshold with 91.3% success rate, GRU model 300 epoch and temperature sampling method best result with 93.8% success rate at 0.5 threshold, Encoder-Decoder and Attention model gives the best result with a success rate of 95.7% at 500 epochs and 0.2 thresholds of the temperature sampling method. At the end of the study, it is seen that the language model created by using encoder-decoder and attention architecture has higher success in producing meaningful words and creating longer word groups than other models (LSTM, GRU). It is seen that the character-based models are successful in producing words and word groups in accordance with Turkish grammar rules.