Published Paper Details:

There is an influx of digitally generated content and user opinions over several decades and has brought major changes to the information environment. Online platforms such as movie review sites, e-commerce portals, and social networking services produce a large amount of user-generated content every year. This includes customer feedback, reviews of products and services, and general social discussions. Together, this information forms an important source of sentiment data that supports business decisions, political communication, and social research. This paper proposes a model SentiBERT with sentiment intelligence framework that uses three transformer architectures--BERT, RoBERTa, and DistilBERT--against classical TF-IDF-based baselines on the IMDB Large Movie Review Dataset, while evaluating across Amazon product and Yelp restaurant review datasets considering cross-domain transfer protocol. SentiBERT generates three remarkable outputs which distinctively elevates this model from rest. First, we design a multi-method Explainable AI (XAI) pipeline that unifies Local Interpretable Model-agnostic Explanations (LIME), BertViz attention-weight visualization, and GPT-2-grounded for natural language explanations-- In this approach, the generated explanation is based on measurable evidence obtained from LIME rather than relying only on unrestricted language model outputs, which helps reduce the chances of imagined results. Second, we implement a hybrid aspect-based sentiment analysis (ABSA) module that integrates spaCy noun-phrase extraction with BERT context-window scoring across six film-specific aspect categories. Third, vocabulary overlap is examined to understand transfer difficulty. For this purpose, Jaccard similarity is used as a reference measure. Additionally, words associated with model failure are identified across target domains using LIME. The experimental results are validated against reproducible JSON metric files which demonstrates RoBERTa achieves the highest test-set accuracy of 91.70% and AUC-ROC of 0.972, representing a 32% reduction in error rate relative to the best classical baseline (Linear SVM: 87.62%). Interestingly, the accuracy drop for zero-shot cross-domain transfer is only 0.60% and 0.40% on Amazon and Yelp, respectively, and contradicts the notion of domain fragility for transformer-based sentiment classifiers. Under text-casing perturbation, BERT achieves 90.0% accuracy, whereas it drops to 70.0% when the test set is heavily corrupted with 20% of typing errors (typo injection). The biggest challenge BERT faces is the identification of subword tokenisation boundaries. The results shed light on the advantages and disadvantages of applying sentiment analysis based on transformers to real-world applications.