Aspect-based sentiment analysis (ABSA) identifies sentiments associated to specific features, giving detailed insights into opinions. Although generative AI models have recently achieved promising results in this area, they often face issues such as inconsistent sentiment polarity and repetitive words, which can affect reliability. This research addresses these limitations by refining key model parameters, applying constrained decoding, and using advanced techniques to improve ABSA performance. Our approach focuses on capturing aspects, opinions, categories, and sentiments more consistently and accurately. Experiments on ABSA datasets, including Aspect Category Opinion Sentiment (ACOS) and Aspect Sentiment Triplet Extraction (ASTE), show improved performance of model. This work enhances generative ABSA model reliability using F1-score, precision, and recall evaluations, supported by constrained decoding and attention regularization, advancing sentiment analysis for various applications.