Performance analysis and summarization of financial reports with natural language processing approaches

In this thesis, summarization and categorization tasks for processing Turkish financial texts are addressed. The main objective of the study is to develop models that improve the summarization and categorization of financial texts while preserving their contextual meaning. The hypotheses are that Transformer-based models can produce summaries with high contextual accuracy and classification models can categorize texts with high accuracy rates. In the study, two separate datasets containing 8,524 texts for classification and 10,288 texts for summarization were used. The datasets were compiled from financial reports, analyses and market news. For summarization, a Cosmos LLaMA Instruct model was customized; for classification, a deep learning-based model optimized with different hyperparameters was developed. The performance of the models is extensively evaluated using contextual and surface metrics. The results show that the summarization model achieves a BERT Score of 66.41% and a semantic similarity score of 63.84% in terms of contextual accuracy. The classification model achieved 83.52% accuracy, 85.53% F_1score and 83.65% precision. The results show that the limitations of Turkish being a low-resource language can be overcome and the developed models can make significant contributions to financial analytical processes. The study also demonstrates the effectiveness of memory-friendly methods with low hardware requirements. This thesis makes a significant contribution to the development of Turkish DDI processes and suggests new research areas for future work.