Abstract/Summary

This thesis explores sentiment analysis in Glassdoor employee reviews, focusing on both English and multilingual contexts. By applying Natural Language Processing (NLP) techniques, we provide a comprehensive review of sentiment analysis in finance, its impact on financial outcomes, and the challenges associated with multilingual sentiment classification. First, our research investigates the practical deployment and evaluation of various NLP models, ranging from lexicon-based approaches to machine learning models, and to state-of-the-art pre-trained language models. By comparing the performance of various sentiment analysis methods, we demonstrate the superiority of advanced models that consider contextual information. These models can substantially enhance sentiment analysis accuracy when compared to traditional dictionary-based approaches. Second, our exploration of multilingual sentiment analysis reveals the impact of translation on sentiment classification. We observe the influence of translation on sentiment misclassification rates, with text attributes playing a more significant role than the quality of translation itself. This suggests that even if the translation quality is high, the sentiment expression might be lost during the translation process, thereby driving the sentiment misclassification rate on translated texts. Additionally, our findings highlight the benefits of zero-shot transfer, demonstrating the effectiveness of fine-tuning multilingual language models when labelled multilingual data is limited. Third, in examining the correlation between employee satisfaction and stock returns, we reveal the predictive power of sentiment measures derived from employee reviews. In our study, we demonstrate that sentiment measures, particularly when derived from BERT, a highly advanced and widely acclaimed language model, effectively predict significant increases in stock returns over both short and long-term periods. BERT, which stands for Bidirectional Encoder Representations from Transformers, is a cutting-edge natural language processing model known for its exceptional contextual understanding of text. Our research also sheds light on the combined influence of employee satisfaction and employee-related costs on stock returns within specific industries. By considering these factors together, we expand the understanding of the complexities underlying the relationship between employee sentiment and financial performance. This thesis makes a significant contribution to the existing literature. It stands out as the first comprehensive study to undertake a rigorous comparison of 31 sentiment analysis methods, employing the extensive dataset of Glassdoor employee reviews. Moreover, this thesis delves into unexplored territory by examining multilingual sentiment analysis within the context of finance. This novel exploration unveils the challenges and implications associated with sentiment classification in a linguistically diverse landscape. This thesis is also distinctive in its pioneering application of BERT to study the correlation between employee satisfaction and stock returns. The findings from this thesis establish a strong basis for future studies in the areas of sentiment analysis, employee satisfaction, and their impact on finance. The practical significance of our work extends to investors and organisations, enabling them to make informed, data-driven decisions that promote employee well-being and enhance corporate performance.