|Friday, February 24|
|CS07 Surveys and Sentiment Analysis||
Fri, Feb 24, 11:00 AM - 12:30 PM
City Terrace 7
Sentiment Analysis of Brand Social Mentions: The Polarity Classification and Beyond (303333)Danielle Leigh Boree, Johnson & Johnson Vision Care, Inc.
Terri Henderson, Johnson & Johnson Vision Care, Inc.
*Jin Su, Johnson & Johnson Vision Care, Inc.
Keywords: Sentiment Analysis, Opinion Mining, Text Mining, Polarity Classification
Consumer opinions are vital to the product development and improvement process in most industries. Social Media mentions and online reviews can provide tremendous insights for a company to monitor the perceived performance of its products versus that of its competitors. This presentation will describe a real-world application of utilizing sentiment analysis to direct product refinement and development.
Over 20,000 online reviews from the past five years and approximately 50,000 social media mentions since 2015 have been collected for four major contact lens manufacturers. Natural language processing (NLP) tools were utilized to assign a sentiment score to each text both overall and for a series of nine pre-defined product attributes (e.g. Price, Quality, Comfort, Vision, etc...). Both rule-based and statistical sentiment models were tested in the NLP tools; the rule-based sentiment analysis model was ultimately selected to classify sentiment polarity. Approximately 80% of the online reviews were able to be classified using this model and the misclassification rate was below 10% when compared with the consumer ratings.
After the sentiment classification, two applications were explored to provide further business intelligence: (1) text mining algorithms, such as Latent Semantic Indexing, were utilized to identify topics within attribute segments which represented the advantages and disadvantages for each brand compared to its competitors, and (2) a trending report was created to track the trend of sentiment scores over time potentially due to market or product changes.
Several challenges will be discussed in this presentation, such as accurately classifying sentiment when the consumer mentions more than one product, filtering out irrelevant text addressed to the website instead of product, and classifying the sentiment of specific product features. Solutions are proposed in this presentation to tackle each challenge.