The COVID-19 pandemic has significantly impacted global health and society, leading to a surge of discourse on social media platforms, particularly Twitter. Analyzing sentiments expressed in these tweets is crucial for understanding public emotions and attitudes during this unprecedented crisis. This paper presents a novel approach for sentiment analysis that combines hybrid contextual embeddings, utilizing models such as ELMo and BERT, with attention mechanisms to enhance the detection of emotional responses in COVID-19 tweets. The dataset consists of 5,000 tweets collected using the hashtag #COVID19 from March to May 2020. Tweets were manually labeled into five sentiment categories for analysis. The hybrid embeddings capture nuanced contextual information, while the attention mechanism selectively focuses on critical keywords and phrases indicative of sentiments. We evaluate our model on a COVID-19 tweet dataset, achieving a micro-averaged accuracy of 85.3% and an AUC score of 0.91, outperforming both ELMo (82.5%, 0.87) and BERT (84.0%, 0.89) baselines. Our findings demonstrate the efficacy of contextual embeddings and attention mechanisms in enhancing sentiment detection, providing valuable insights into public sentiment during the pandemic. This research contributes to the growing field of sentiment analysis and highlights the importance of advanced techniques in understanding social media discourse.