This validation study investigates the application of sentiment analysis to sports-related textual data. While the categorization of individual tweets remains challenging, the sentiment of large sets of tweets can be categorized with a sufficient accuracy by simple tools. Qualitative observations suggest that the analysis of sports related-content contains its own difficulties. The average tweet length in our database was shorter compared to tweets including common hashtags from other domains (football: 13.3 words; politics: 20.8 words; finance: 18.0 words) making classification more challenging. Moreover, we are not aware of any lexicon that specifically considers the connotation of football-related vocabulary. .Accuracies with regard to individual tweets were 61.0% for LIWC, 63.6% for QDAP, 62.6% for SN and 67.4% for COMB. Accuracies are significantly higher than 50% according to one-sided binomial tests (p < 0.001 for all four), but still far from satisfactory given that 50% would be expected from guessing. However, in most real-world applications the number of tweets is large and thus set accuracy is more meaningful. The tools are capable of classifying the polarity of realistic sets of tweets (e.g., 1000 tweets, with a proportion of 60% having the same polarity) with an accuracy of more than 95%, thus being accurate enough in real-world applications.
© Copyright 2020 spinfortec2020digital: 13. Symposium der dvs-Sektion "Sportinformatik und Sporttechnologie" vom 24.-25. September 2020 in Bayreuth. Veröffentlicht von Institut für Sportwissenschaft. Alle Rechte vorbehalten.