International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 13 Issue: 05 | May 2026
p-ISSN: 2395-0072
www.irjet.net
AN EXPLAINABLE TRANSFORMER-ENHANCED NEURAL FRAMEWORK FOR ASPECT-BASED OPINION MINING USING HYBRID FEATURE EXTRACTION 1 Shubha S,2 S. Shubhakar, 3 Varsha Jayaprakasha 1Associate Professor, Government First Grade College, Malleshwaram, Bangalore, India
2Digital Associate Engineer, Sonata Software Solutions Limited, Global Village, Mysore Road, Bangalore 3Backend Developer, Canibuild, Australia
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Opinion mining has become a critical component
in understanding user-generated content across domains such as tourism, e-commerce and social media. This paper proposes a Transformer-Enhanced Neural Learning-Based Opinion Feature Extraction (TE-NLOFE) model that extends traditional neural approaches by integrating contextual embeddings, attention mechanisms and ontology-based linguistic mapping. Unlike earlier models, the proposed framework captures contextual dependencies using BERT [1] and introduces explainability through attention-based interpretation [8]. Experimental evaluation on benchmark datasets including TripAdvisor, Amazon Reviews and SemEval demonstrates improved accuracy, reduced false positive rate and enhanced interpretability compared to baseline models [10], [11]. The results validate the effectiveness of combining deep learning with linguistic ontology for fine-grained sentiment analysis.
Key Words:
Neural Networks, Feature Extraction, Explainable Artificial Intelligence (XAI), Natural Language Processing (NLP), Text Mining, OntologyBased Analysis, Hybrid Models, Contextual Embeddings.
1. INTRODUCTION With the exponential growth of online reviews, opinion mining has become essential for extracting meaningful insights. Earlier models, including the Neural Learning Based Opinion Feature Extraction (NLOFE), relied on neural networks and linguistic mapping but lacked contextual understanding and scalability [13]. Recent advancements in transformer architectures such as BERT [1] and the foundational work on attention mechanisms [2] have revolutionized natural language processing by capturing contextual semantics more effectively. Traditional sentiment analysis approaches focused on feature extraction and classification [3], [4], but failed to capture aspect-level dependencies. Recent studies in aspectbased sentiment analysis (ABSA) have attempted to address this limitation [6], [7].
This paper proposes an enhanced framework that:
Integrates transformer-based embeddings Performs aspect-level sentiment classification Introduces explainability mechanisms using SHAP and LIME [8], [9] Extends prior work in neural opinion mining [13]
2. LITERATURE REVIEW Opinion mining and sentiment analysis have evolved significantly over the past two decades, transitioning from traditional rule-based systems to advanced deep learning and transformer-based architectures. Early research in sentiment analysis primarily relied on lexicon-based and statistical approaches, where sentiment polarity was determined using predefined dictionaries and linguistic rules. These methods were simple and interpretable but often failed to capture contextual meaning, sarcasm, and domain-specific variations [3]. Subsequently, machine learning techniques such as Support Vector Machines (SVM), Naïve Bayes, and logistic regression were introduced to improve classification performance. These approaches leveraged feature engineering techniques includingn-grams, part-of-speech tagging, and term frequency–inverse document frequency (TF-IDF). While these models improved generalization, they were still limited by their dependence on manual feature extraction and inability to capture long-range dependencies in text [3], [4]. The emergence of deep learning marked a major shift in sentiment analysis research. Neural architectures such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Long Short-Term Memory (LSTM) networks enabled automatic feature extraction and improved performance in sequence modeling tasks. These models were particularly effective in capturing syntactic and semantic relationships within text. However, despite their advantages, deep learning models still faced challenges in handling long-range dependencies and contextual ambiguity due to sequential processing limitations [4]. A significant breakthrough occurred with the introduction of transformer architectures, particularly BERT and related models. Transformers utilize self-attention mechanisms to
© 2026, IRJET
|
Impact Factor value: 8.315
|
ISO 9001:2008 Certified Journal
|
Page 938