Fine-tuning isn't dead yet, at least not for Turkish sentiment analysis.
Researchers compared three approaches on a Turkish e-commerce review dataset: classical machine learning, fine-tuned pretrained language models, and prompted large language models in zero-shot settings. Fine-tuned BERTurk models came out on top across all three sentiment classes — negative, neutral, and positive. Prompted LLMs were competitive in simple positive-negative binary tasks, but fell apart when a neutral category was added, tending to collapse nuanced reviews into one of the two poles.
The neutral class is the real story here. It's easy to build a benchmark that flatters LLMs by stripping out ambiguity, and this study explicitly resists that. The finding matters because e-commerce and customer feedback analysis — the places where sentiment tools actually get deployed — are full of tepid, mixed, or hedged opinions that don't map neatly onto thumbs-up or thumbs-down.
The results won't surprise anyone who has watched LLM benchmarks get gamed by convenient test set choices. Zero-shot prompting is cheaper and faster than fine-tuning, which is why the industry keeps rooting for it to win — but for low-resource languages like Turkish, the gap remains real. Until prompted models learn to sit with ambivalence, fine-tuning holds its ground.