Companies using sentiment analysis could be getting a distinctly female view of the world, according to latest research from statisticians.
Using speedy software to detect customer opinions on products and services has been a standard business technique for over a half a decade.
But new research from the University of Wolverhampton shows that sentiment analysis over-represents the opinions of females because they generally express sentiment more clearly.
Using UK TripAdvisor reviews of hotels and restaurants (chosen as product types important to both men and women), Professor of Data Science Mike Thelwall contrasted the accuracy of lexical sentiment analysis for males and females. Lexical sentiment analysis identifies sentiment-related terms, such as ‘happy’, ‘excellent’ and ‘dirty’ and applies a set of rules to guess the sentiment of the review.
Women reviewers used more words such as ‘lovely’, ‘delicious’ and ‘amazing’; while men tended to use more common factual words such as ‘location’, ‘building’ and ‘beer’. Professor Thelwall said: “This is an important difference because sentiment analysis programs rely on people expressing opinions in predictable ways.”
Male sentiment was harder to detect because it was less explicit, with women using more direct language to express the same sentiment.
“When comparing opinions for product aspects that appeal differently to men and woman, female sentiments are likely to be overrepresented, biasing the results,” said Professor Thelwall.
“Companies and political parties that rely on sentiment analysis are getting a slightly female-tinged view of the world.”
He added: “This is the first evidence that lexical sentiment analysis is less able to detect the opinions of one gender than another. Care should be taken when drawing conclusions about gender differences from automatic sentiment analysis data.”
To conduct the research, Professor Thelwall gathered a large collection of reviews with ratings and author genders and then assessed for accuracy a sentiment analysis programme on appropriate subsamples written by males or females. Accuracy was found to be significantly lower for male-authored reviews than for female-authored reviews for restaurants and all types of hotels tested.
Notes to Editors:
A copy of the full research paper, which has been published on research site Emerald Insight, is available on request.