Sentiment analysis is an important tool in understanding users of the Web through the cues they leave while communicating and providing feedback. Sentiment classification models may have lower robustness because they detect irrelevant patterns instead of sentiment-bearing words related to the target aspects. We evaluated the robustness of the Hybrid Approach for the Aspect-Based Sentiment Analysis++ (HAABSA++) model. We first generated an Aspect Robustness Test Set (ARTS) through the augmentation of the original test set with different augmentation techniques: RevTgt, RevNon, and AddDiff. These techniques modify the target and non-target aspects to test whether the model still makes correct predictions. Aspect robustness is evaluated using the Aspect Robustness Score (ARS). In addition, we investigated the improvement of ARS through adversarial training by applying the three data augmentation methods to the training set. We find that the robustness of the HAABSA++ model as measured by ARS is average when we compare the HAABSA++ model with other models found in the literature. We also find that performing adversarial training improves the robustness of the model as measured by ARS. However, this improvement comes at the cost of a lower accuracy of the HAABSA++ model for the original test set instances.