Semantic relations-based summarizing of scientific articles

Text summarization is a very important method for researchers to grasp a document as quickly as possible. The document is summarized with approaches based on the removal of important sentences or rewriting the text. With the development of deep learning technologies, documents are summarized by extracting semantic relationships between words starting from the beginning of the text, just as if they were produced by a human. This approach is known as abstractive summarization. Abstractive summarization makes it difficult to summarize scientific articles in terms of preserving the semantic content between the words, as scientific articles are in the category of long documents due to their structure and contain up-to-date terminological words that are constantly innovating. In this thesis, two different models are proposed to summarize scientific articles. In the proposed text summarization model with graph-based method, the summarization of articles related to computer science is performed. In this model, there is no token limit for the document. Summary generation was performed in the model by combining the entity names and relations extracted from scientific articles with the document content. The proposed model has shown superior performance in summarizing scientific articles compared to the baseline models. In other studies, scientific articles were summarized with as few but important words as possible. A novel model was developed by extracting the most important words in the document. When the proposed method was compared with the basic models, successful summaries were produced with as few and important words as possible, such as summaries written by the author.