Efficient Multi-Class Analysis of Consumer Complaints Using Frozen MiniLM Embeddings and Neural Networks

Text classification is a critical task in domains generating large volumes of unstructured text, such as finance, healthcare, and consumer services. However, accurately classifying such data remains challenging due to its noisy, imbalanced, and context-dependent nature. While pre-trained language models have improved general text classification, their direct application often overlooks domain-specific cues and sentiment patterns that are important for nuanced understanding. In this study, we propose a novel framework that extends the MiniLM language model by integrating domain-relevant cues and sentiment features with textual embeddings. This integration allows the model to capture both semantic richness and domain-specific patterns, enhancing reliability and interpretability. Comparative experiments against baselines including TF-IDF + Logistic Regression, Word2Vec + Logistic Regression, TF-IDF + Naïve Bayes, and Word2Vec + Naïve Bayes shows that the proposed approach consistently outperforms traditional methods, achieving an accuracy of 0.8653, precision of 0.8697, recall of 0.8653, F1-score of 0.8668, Cohen’s Kappa of 0.7862, and MCC of 0.7870. Ablation studies further demonstrate the critical role of cues and sentiment features in improving performance. These findings indicate that combining pre-trained embeddings with carefully selected domain features offers a more robust and context-aware solution for text classification, establishing a foundation for future work integrating transformer-based models with explainable AI techniques in domain-specific applications.

@article{Gopinathan2025,
  title     = {Efficient Multi-Class Analysis of Consumer Complaints Using Frozen MiniLM Embeddings and Neural Networks},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {16},
  number    = {12},
  year      = {2025},
  publisher = {The Science and Information Organization},
  author    = {Sri Vishnu Gopinathan and Muhammad Faraz Manzoor},
  doi       = {10.14569/IJACSA.2025.0161202},
  url       = {https://doi.org/10.14569/IJACSA.2025.0161202}
}