Lak_Parisa.pdf (1.68 MB)

The impact of sentiment analysis on decision outcomes - an empirical investigation

Download (1.68 MB)
thesis
posted on 08.06.2021, 13:55 by Parisa Lak
A typical trade-off in decision-making is between the cost of acquiring information and the decline in decision quality caused by insufficient information. Consumers regularly face this trade-off in purchase decisions. Online product/service reviews serve as sources of product/service related information. Meanwhile, modern technology has led to an abundance of such content, which makes it prohibitively costly (if possible at all) to exhaust all available information. Consumers need to decide what subset of available information to use. Star ratings are excellent cues for this decision as they provide a quick indication of the tone of a review. However there are cases where such ratings are not available or detailed enough. Sentiment analysis - text analytic techniques that automatically detect the polarity of text - can help in these situations with more refined analysis. This study was performed in two interrelated phases. In the first phase the potential impact of Sentiment Scores (sentiment analysis outcomes) was investigated through a comparison between these scores with an already established numerical rating denoted as star ratings in three different domains. The results show that sentiment scores tend to fall into neutral areas and are not able to detect extremes that were reported to be more beneficial for information acquisition purposes. As a result, to use the current sentiment analysis results as a substitute for star ratings, a partial linear filter was applied to sentiment analysis results in a way to highlight the subtle differences away from the "neutral zone". In the second phase, the impact of the extended version of sentiment scores on decision outcomes was examined through a controlled experiment. The examined decision was a purchase decision and the information provided was pages of reviews annotated with extended sentiment scores on each paragraph. Human subjects were used in the experiment and controlled data gathering sessions was designed. Results suggest that female consumers may use sentiment scores on review documents without other comparison aids to increase their confidence level in their purchase decisions.

History

Language

eng

Degree

Master of Management Science

Program

MBA in Management of Technology and Innovation

Granting Institution

Ryerson University

LAC Thesis Type

Thesis