<![CDATA[Story generator plays an important role to help story writers generate story ideas, even initial concepts. Usage of Keras Tokenizer as well as word embedding model requires relatively slower model training speed in order to execute hundreds of training iterations. In this research, we propose design method to create a word-level Indonesian language story generator by implementing Markov Chain model and Bidirectional GRU, which is able to generate quality text as good as the outputs of word embedding models, while having faster model training speed. The performance of Markov Chain-BiGRU model was compared with the performance of word-level BiGRU model and character-level GRU model. The first stage of model evaluation was done by comparing each model’s loss value and model training speed; the second stage was done by giving survey to 33 assessors; while the third stage was done by comparing model’s performance with model from related work. The proposed Indonesian story generator succeeded on increasing the model training speed by 66.38% from related work’s model, as well as producing better-quality text compared to outputs from conventional neural-based and word embedding models.]]>
Copyrights © 2023
