Fine-tuning Turkish LLM With Human-Annotated Paraphrase Corpus

Paraphrasing is commonly used in daily language, and both theoretical and applied research on paraphrasing exists. In natural language processing, paraphrasing tasks have garnered significant interest and have been the focus of much research. With this interest, this study specifically focuses on Turkish paraphrase generation, an area with limited research. In the first phase of the study, we aimed to create a Turkish paraphrase corpus annotated by twice. Next, we fine-tuned the corpus on the Turkish large language model, TURNA, and finally we obtained the best results in NLU mode with 75.9 Rouge1, 61.4 Rouge2, 73.5 RougeL, 49.6 BLEU, 72.5 METEOR scores.