Fig. 4

Main indicators of the ACR method. Regression lines reflect ordinary least squares regression, with \(95\%\) confidence intervals (estimation by bootstrapping). In panels b and g, colored areas and white triangles reflect KDEs and means, respectively. a Mean text similarity and sentiment score across baselines and story period. Error bars reflect \(95\%\) confidence intervals. Both measures showed the expected pattern of a peak during the story period. b Mean text similarity was significantly higher for stories with four baselines, indicating that the ACR method is more reliable when more baselines are available. c A strong association between the number of matching tweets and mean text similarity is a signature of occasional tweet overload. d A weak association between mean text similarity and sentiment score indicates that the more reliably a story is measured, the more negative the tweets are. e KDE of the mean correlation between subqueries. A majority of the subqueries are highly correlated. f Association between mean correlation between subqueries and text similarity; the more the subqueries converge, the better the story is measured. g A text similarity comparison of the ACR-retrieved and original tweets indicates that ACR-retrieved may outperform original tweets. h An association between text similarity of the ACR-retrieved and fact-checked tweets indicates that the reliability of the ACR depends upon a good training set