Employee Satisfaction Mining: A Deep Learning Approach Using LLMs for Real-Time Sentiment Monitoring

The levels of employee satisfaction hold a fundamental capacity for impacting performance within an organization, being able to affect productivity, retention, and even innovation. An age-old way to gauge employee sentiment was through surveys and interviews, where there is usually lag time in receiving results, where results are highly localized from view points within certain disciplines, and from inherent bias. In our work, we propose a framework for real-time sentiment monitoring that employs fine-tuned Large Language Models (RoBERTa) to capture employee sentiments from internal communication platforms like Slack and Microsoft Teams. Moving forward with our method, after secure data gathering, the data is anonymized, preprocessed, and subjected to sentiment classification, considering three sentiment levels: Positive, Neutral, or Negative. Real-time sentiment trends are then showcased on a live dashboard, enabling intervention when necessary by the human resource departments. The model attained an F1-score of 90.8% subsequent to fine-tuning on a hybrid dataset amalgamated from organizational communication data and public sentiment corpora, much outperforming conventional models such as SVMs, LSTMs, and naive Bayes classifiers. This model facilitates nearly real-time monitoring with a latency of fewer than two seconds per message and upholds data privacy norms including GDPR. The framework has been conceptualized in-concert with ethical considerations being central to it, including bias audits and data usage based on consent. In conclusion, the research does show the practical viability of LLMs in altering HR analytics by providing timely, scalable, and context-aware insights into sentiment.