Deep Learning Based Abstractive Turkish News Summarization

Abstract

With the increase of knowledge, there is a need for summarization systems that will direct the person to the area they are interested in without any waste of time. In this work, Turkish news headlines have been predicted by using encoder-decoder model from deep learning methods. Abstraction based text summarization method has been used during the generation of headlines. The system has been trained with recurrent neural networks by developing encoder-decoder model. The word embeddings of the words in news texts have been generated by using FastText that is very commonly used model in the literature recently. The system has been tested separately by training the first sentence, first two sentences and full-text of each news. The success of the system is measured by ROUGE score and semantic similarity score. According to the experimental results, it has been observed the model trained with full-text of news outperforms among the other models.