Text summarization using machine learning

The abundance of information sources and the size of the documents containing information cause the process of obtaining information from the text to take a lot of time. Developments in information technologies have led to the search for different methods that can be used to access information from text. Automatic text summarization systems have been developed in order to access information from text. Automated text summarization systems take a document as input and output important parts of the document in a more concise, understandable way. Automatic text summarization reduces the time spent retrieving information from long texts and allows more documents to be examined in a limited time period. Automatic summarization, which is one of the fields of study of the Natural Language Processing branch of science, has been carried out with various statistical and deep learning-based methods since 1958. After the use of Transformer architecture in natural language processing studies, large language models have emerged and it has been observed that these language models achieve very successful results in text summarization processes. Computers with powerful hardware are needed to retrain large language models, and this is a very costly process. The ability to use these language models by ordinary users is possible with transfer learning methods. In this thesis study, a comment-based automatic text summarization study was carried out in Turkish. For this study, a total of 200 scientific articles were collected through Dergipark, an online publishing tool for scientific journals. A Turkish article data set was created with these scientific articles collected from journals published in the fields of educational sciences, information technologies, economics and business science. The data set includes the entire text content of the scientific article, its abstract, field of science, name of the article and keywords. A new language model was obtained by fine-tuning the PEGASUS language model, which was pre-trained on an English dataset from large language models, with the created Turkish article dataset. With the language model obtained, a comment-based text summarization study was carried out on the Turkish article data set. In order for automatic text summarization to be performed, the text must be divided into tokens. In this study, a tokenizer specific to the data set was created and this tokenizer was used in the comment-based text summarization process. The created summaries were compared with the summaries obtained using the language model'sown tokenizer. The comparison process was carried out by using the summaries obtained after automatic text summarization and ROUGE metrics, and by examining their semantic integrity, considering the summary written by the article author as the reference. It was observed that the summarization study conducted with the tokenizer created specifically for the data set yielded results more similar to the content of the article, although the ROUGE values were lower. For the PEGASUS language model, which was pre-trained on an English data set, fine-tuning with a Turkish data set specific to a specific purpose has not been found in the literature. For this reason, the obtained results could not be compared with different studies. It is thought that this study will form a basis for future studies in Turkish language with the PEGASUS language model.