Google Pathways Language Model Powered Analysis of Telecom Transcripts for Sentiment Detection

Abstract

The rapid expansion of the telecommunications sector has generated vast volumes of customer–agent interaction data, rendering manual sentiment analysis impractical. This research proposes a robust and scalable framework for telecom transcript sentiment detection by integrating advanced natural language processing (NLP) techniques, transformer-based embeddings, and ensemble machine learning models. The framework begins with comprehensive data preprocessing, including text cleaning, tokenization, stopword removal, and lemmatization. This is followed by exploratory data analysis using word clouds, document length distributions, part-of-speech (POS) tagging, and bigram frequency analysis to uncover underlying textual patterns. To capture rich contextual semantics, Google PaLM-like embeddings are generated using Hugging Face transformer models. Class imbalance is effectively handled Random Under Sampler to ensure uniform representation across sentiment classes. Multiple machine learning models Logistic Regression Classifier (LRC), Decision Tree Classifier (DTC), Extra Trees Classifier (ETC), Boosted Rules Classifier (BRC), and a custom FIGS (Fast Interpretable Greedy-Tree Sums) ensemble classifier are trained and evaluated. The FIGS model aggregates predictions from base learners to improve accuracy, robustness, and generalization. The proposed system supports real-time sentiment prediction, model persistence, and visualization of performance metrics, offering interpretable insights for telecom operations. Evaluation metrics, including accuracy, precision, recall, and F1-score, demonstrate the effectiveness of the framework. The system provides an efficient, scalable, and interpretable solution for automated sentiment detection in telecom transcripts, enabling service providers to enhance customer experience, monitor agent performance, and make informed, data-driven operational decisions.