Differentiated Sentiment Analysis

Differentiated Sentiment Analysis

Practitioners and researchers alike increasingly use social media messages as an additional source of information e.g., to analyze stock price movements, assess online reputation, or predict election results. When applying automated sentiment analysis, IS researchers have predominantly focused on measuring the bipolar valence dimension of positive-negative emotionality (e.g., Sprenger et al., 2014). However, the undifferentiated dimensional approach implies a lower degree of specificity which can only be overcome through the assessment of distinct emotional states (Ekkekakis, 2013). Emotion researchers generally distinguish between emotions as distinct-states or dimensions. In a critical reconsideration of these different approaches Ekkekakis (2013) integrated both understandings into one model of the hierarchical structure of the affective domain (figure 1 depicts the adapted version of Ekkekakis' (2013) framework).

Figure 1: Adapted Version of the Model of the Hierarchical Structure of the Affective Domain

Based on the concept of seven different emotion states in the model, we developed an operationalization of seven different emotions (affection, happiness, satisfaction, fear, anger, depression, and contempt) (see table 1) which we applied to the existing set of emotion words from the lexicon of the established sentiment tool “SentiStrength 2” (Thelwall et al., 2012). The developed dictionary thereby exceeds existing sentiment tools by enabling to analyze differential emotions while simultaneously considering the strength of the emotions and the exclusiveness of emotional states.

Table 1: Overview of the Seven Different Emotions and their Operationalization

In a first study, we applied the dictionary to investigate the connection between the differential emotions and stock movements where we analyzed approximately 5.5 million Twitter messages on 33 S&P 100 companies and their respective NYSE stock prices from Yahoo!Finance over a period of three months. The results generally support the assumption of the necessity of considering a more differentiated sentiment. Moreover, comparing positive and negative valence, we find that only the average negative emotionality strength has a significant connection with company-specific stock price movements. The emotion specific analysis reveals that an increase in depression (see figure 2) and happiness strength is associated with a significant decrease in company-specific stock prices.

Figure 2: Juxtaposition of the Depression Sentiment Score and a Company Specific Stock Price Movement Over a Three Month Period

In a second study, we measured the strength of the seven different emotions within 532,363 tweets explicitly addressing one of 641 accounts from 33 S&P 100 companies. Thereby, we differentiate between Branding (Advertising, Public Relations, and News Releases), Sales, Customer Service & Support, Product Development, and Human Resources (Jobs & Careers and Academy & University) accounts. Our results (see table 2) show that the accounts also differ regarding the emotions expressed about them. Customer service accounts are subject to most angry, depressed and contemptuous messages (with contempt as the strongest differentiating emotion), while users express strong happiness towards the branding and especially human resources accounts. Satisfaction, however, seems to be a highly specific emotion which only differs significantly between low-level account categories, whereas expressions of fear and affection do not vary between account types.

Table 2: Results of the A-Posteriori ANOVA Account Comparisons Regarding the Average Emotion Strength

The dictionary is currently being revised and will soon be available for download again

References

Ekkekakis, P. (2013). The Measurement of Affect, Mood, and Emotion: A Guide for Health-Behavioral Research. Cambridge University Press.

Sprenger, T. O., A. Tumasjan, P. G. Sandner and I. M. Welpe (2014b). Tweets and trades: The information content of stock microblogs. European Financial Management, 20 (5), 926–957.

Thelwall, M., K. Buckley and G. Paltoglou (2012). Sentiment Strength Detection for the Social Web. Journal of the American Society for Information Science and Technology, 63 (1), 163-173.