Relational document classification with deep learning methods

Text classification, used to extract meaningful information from vast and complex pools of text and to categorically organize it automatically, is one of the applications of natural language processing (NLP). However, these methods can be insufficient for languages with rich morphological structures and diverse data sources. To address this issue, this study investigates the integration of entity-relation information from Wikidata into texts using Named Entity Recognition (NER) models and pre-trained BERT-based models. Evaluations conducted on Turkish and English text datasets have examined the performance of various machine learning and deep learning classification algorithms. Findings indicate that the addition of entity-relation information significantly enhances the performance of algorithms. Particularly, the integration of relational information has notably improved classification success in the BBC-News and TTC-4900 datasets. For effectively processing the complex structures of Turkish texts, graph-based deep learning techniques have been applied to a dataset obtained from the TRT-News website, achieving a noteworthy accuracy rate. The results demonstrate the efficacy and potential of adding relational information in text classification, contributing to a better understanding of Turkish texts. This research has made significant contributions to the development of relational text classification systems and has established an important reference point for future studies.