Abstract
This article discusses the inconsistencies, inaccuracies and challenges, namely the 'dark side' of sentiment analysis and then demonstrates problems with using sentiment analysis lexicons or dictionaries for estimating sentiment in textual artifacts. Sentiment analysis, an important dimension of natural language processing (NLP), has seen an exponential adoption rate across research and practitioner disciplines. Many interesting developments in NLP methods continue to improve the accuracy of sentiment analysis. However, the plethora of sentiment analysis methods, dictionaries and lexicons, tools, open source code for machine learning based sentiment analysis, and of-the-shelf sentiment analysis solutions have led to a flurry of research and applied solutions without sufficient concern for the limitations, context, and the inaccuracies of sentiment analysis, and the inherent ambiguities associated with the unaddressed sentiment analysis domain challenges. Scant attention is given, especially in applied research and industry usage, to the inherent ambiguities associated with the unanswered questions pertaining to the science of sentiment analysis. This study reviews known issues with sentiment analysis as documented by prior research and then compares the application of multiple of-the-shelf lexicon and dictionary methods to stock market and vaccine tweets. The intention is not in any way to improve the accuracy of sentiment analysis as compared to prior benchmarks but to identify and discuss critical aspects of the dark side and develop a conceptual discussion of the characteristics of the dark side of sentiment analysis. We conclude with notes on conceptual solutions for the dark side of sentiment analysis and point to future strategies that could be used to improve the accuracy of sentiment analysis and understanding. This research will also help align researcher and practitioner expectations to understanding the limits and boundaries of natural language processing based solutions for sentiment analysis and estimation.