A Comparative Study of Qwen and Gemma Large Language Models for Social Media Slang and Emoji Sentiment Analysis for Business Recommender Systems

Sentiment analysis of social media text remains a fundamentally challenging task due to the pervasive use of informal slang, abbreviations, deliberate spelling modifications, and emojis. Traditional lexicons and early machine learning models frequently struggle to decode the nuanced semantics of these modern digital expressions, particularly when distinguishing between genuine sentiment and sarcasm. In this paper, we present a comprehensive comparative study investigating the capabilities of two recent 4-billion-parameter large language models (LLMs), Qwen3.5 (4B) and Gemma3 (4B), in robust social media sentiment analysis. The core contribution of our study is the fine-tuning of these models at a comparable scale using 4-bit NormalFloat (NF4) quantization combined with the Unsloth library, establishing a highly efficient and reproducible training pipeline. We critically evaluate both models in zero-shot and fine-tuned settings on a diverse, emoji-heavy dataset characterized by four complex sentiment classes: Positive, Negative, Neutral, and Sarcastic. Our experimental results demonstrate that while both models show highly capable zero-shot generalization (Qwen achieving 75.95% accuracy and Gemma achieving 73.20%), they exhibit distinct failure modes surrounding sarcastic interpretations. However, alignment via Low-Rank Adaptation (LoRA) dramatically elevates their performance, allowing both to achieve 100% accuracy on the held-out test set while maintaining distinctly competitive inference speeds suitable for real-time applications. We also provide an exhaustive architectural analysis of the models' inference latency, per-class prediction bottlenecks, and training efficiency, highlighting practical considerations and trade-offs for deploying lightweight LLMs in real-world sentiment analysis environments. The strong inference-speed profile of both fine-tuned models makes them directly deployable as opinion-mining modules within business recommender system pipelines, where timely and accurate decoding of customer-expressed sentiment—including sarcasm—translates into richer preference signals and more reliable personalization.