Optimizing Small Transformer-Based Language Models for Multi-Label Sentiment Analysis in Short Texts