TR-SUM: Türkçe için bir metin özetleyici

In today's world of big data, the tracing and obtaining of specific and useful information from stored and textual data is one of the most significant challenge of summarization. Summarization is basically the process of shortening the text. However, retrieving a short summary or only relevant information from large texts is demanding. The summary should include the main purpose, main idea, and important information of the text. The studies for both the extractive and abstractive text summarization models are recently increasing for various languages. However, fewer studies have been carried out for the abstractive text summarization of Turkish language. In addition, there are less collected datasets to be summarized for Turkish language in the literature. Thus, the contribution of this thesis is in two-fold. Firstly, a news dataset is collected for Turkish language. Secondly, this thesis proposes the adaptation and implementation of three abstractive deep neural network models for the Turkish language. These models are (i) Attention Based Seq2Seq Neural Network, (ii) Pointer Generator Seq2Seq Neural Network and (iii) Reinforcement Learning with Seq2Seq Neural Network. All three models are preprocessed with both ConceptNet-Numberbatch word embedding and fastText word embedding. Then, these models are evaluated on the collected Turkish dataset based on the ROUGE-1, ROUGE-2, and ROUGE-L scores. According to the computation experimentation, All ROUGE scores of each neural network model that is studied with fastText word embedding have decent quality. The highest ROUGE scores are acquired by Pointer Generator Seq2Seq Neural with fastText word embedding. This indicates that Pointer Generator Seq2Seq Neural Network is a promising deep neural network model that may produce well-qualified text summaries.