Aviation is one of the predominant sectors that contribute significantly to the global economy. With the advent of technology, this industry is witnessing a paradigm shift towards data-driven approaches. The morale of the airline employees is barely noticed, which causes fatigue and depression. Furthermore, these mental health issues can be active reasons for destructive accidents. In this research, the authors are focused on collecting insightful information on aviation employees from Glassdoor.com. Moreover, the authors focus on analyzing the sentiments of the employees of renowned aviation companies. Primarily, the authors scraped necessary data from Glassdoor.com and created a dataset named JetJobJoy (JJJ). Data quality is measured with the Inter Annotator Agreement (IAA), in which three experts from the concerned domain ensure the credibility of the dataset. An extensive Exploratory Data analysis is performed to extract essential factors from the dataset. The dataset contains several attributes, such as the company’s rating, the job’s pros and cons along with the feedback of the employees regarding their workplace. The feedback comments are furthermore preprocessed properly and fed into numerous sequence-to-sequence and transformer-based architectures. Furthermore, an improvised architecture of ModernBERT has been proposed with a lesser number of encoders that outperforms other state-of-the-art architectures in terms of performance metrics (95.69% F1-score) and sustainability. The model is also utilized to perform on other datasets for detecting cyberbullying and shows promising results. Finally, the authors have undertaken the diligent effort to interpret the model with the LIME Explainable AI model. • Analyze workforce feedback to uncover sentiment trends within the aviation industry post-pandemic. • Develop a decision support model using transformer-based sentiment classification. • Build and validate a large aviation review dataset to support future analytics research. • Enhance transformer performance through architectural improvements and fewer trainable parameters. • Interpret model predictions using explainable methods to support transparent decision-making.
Chakraborty et al. (Sun,) studied this question.