Entity Resolution (ER) is a foundational task in data integration and processing, focused on accurately identifying and merging records that represent the same real-world entities. This study introduces a Transformer-based Denoising Adversarial Variational (DAV) framework to tackle ER challenges using synthetic datasets, such as DBLP-ACM and iTunes-Amazon, which simulate real-world data heterogeneity, noise, and duplication. Our approach leverages a Sentence Embedding Module to capture nuanced semantic information from entity records, transforming them into robust embeddings. These embeddings are further refined through an adversarial autoencoder, converting low-level representations into high-level probability distributions that enrich the entity matching process. A supervised matching model is subsequently employed to classify paired records, evaluating whether they correspond to the same entity. To assess model effectiveness, we benchmark it against leading ER approaches, such as DeepMatcher and DITTO, analyzing accuracy and resilience across synthetic data variations. An explainable AI component is also integrated to reveal key attribute contributions, offering insights to fine-tune model hyperparameters for optimized performance and interpretability.