Comparison of Fine-Tuning and Prompt Strategies for Large Language Models in Implicit Depression Sentiment Analysis

The detection of depression through textual analysis has evolved significantly with the adventof Large Language Models. While traditional methods focused on explicit keyword detection,modern challenges lie in identifying implicit depression, which manifests through metaphors,subtle semantic shifts, and nuanced emotional undercurrents rather than overt declarationsof sadness. This paper presents a comparative study between two primary adaptationparadigms for Large Language Models: parameter-efficient fine-tuning and promptengineering strategies. We evaluate the performance of models such as Llama 2 and GPT-4on a curated dataset of implicit depression indicators derived from social media platforms.The study specifically investigates three prompting techniques—zero-shot, few-shot, andchain-of-thought—against Low-Rank Adaptation fine-tuning. Our results indicate that whilefine-tuning achieves superior F1 scores by adapting the internal representations of the modelto the specific linguistic patterns of the depressive domain, advanced prompting strategieslike chain-of-thought offer a competitive alternative with significantly lower resourceoverhead. We further analyze the trade-offs regarding computational cost, inference latency,and the risk of hallucination. The findings suggest that while fine-tuning remains the goldstandard for accuracy in high-stakes clinical monitoring, prompt engineering provides aviable, scalable solution for preliminary screening and real-time intervention systems.