Turkish Text Summarization with Artificial Intelligence and Deep Learning Methods: Improving the Performance of Pretrained Large Language Models for Turkish through Vectorial Cosine Similarity-Based Content Selection

Özet

This paper aims to present an innovative solution for the Turkish text summarization process by leveraging artificial intelligence and deep learning-based large language models (LLMs), specifically focusing on the use of pre-trained language models. In this context, the primary objective of this study is to adapt pre-trained models for the task of Turkish text summarization, thereby proposing a cost-efficient and performance-oriented approach that eliminates the need for extensive training. The theoretical foundation of this research is based on the numerical representation of the distributional regularities of language, positing that leveraging the similarity ratios of these regularities can enhance the structured output generation of generative language models. In this regard, it is expected that large language models will generate more structured and coherent outputs by utilizing the contextual similarity between sentence vectors in a vector space where each word is represented as a vector. This dichotomic approach aims to improve the consistency of generative outputs by increasing the regularity of input vectors through a separate process that computes the similarities between them. In this study, a cosine similarity-based content selection strategy for sentence vectors in large language models is proposed for Turkish text summarization. The objective is to select the most meaningful sentences, providing the model with more targeted and summarization-friendly input, thereby enhancing summarization performance while reducing computational costs. The study investigates the improvement of the T5 model’s performance by feeding it with sentences that exhibit the highest vectorial similarity. The evaluation is conducted using ROUGE-1, ROUGE-2, and ROUGE-L metrics. The results indicate that selecting sentences with high vectorial cosine similarity contributes to a significant improvement in summarization quality. Furthermore, statistical analyses using t-tests demonstrate that this method enhances summarization performance in certain cases.