Fig 1.
Diagram showing the structure resulting from the journalists’ annotation work. Rumour stories, represented by squares, can be one of true (green), false (red), or unverified (orange). Each of the rumour stories has a number of rumour threads associated with it (black lines). When a story is true or false, the journalists also picked, where available, one tweet as the resolving tweet within the story’s timeline.
Table 1.
Outcome of the annotation of rumours.
Fig 2.
Example of rumourous conversation.
Example of a conversation generated by a rumourous tweet.
Fig 3.
Annotation scheme for rumourous social media conversations.
Fig 4.
Example of annotation of rumour type, as well as how we determine support towards the rumour.
Table 2.
Inter-annotator agreement values for different features and tweet types.
Fig 5.
Timelines of rumour lifecycles.
Rumour timelines showing the lifecycle of rumours that start as unverified stories (orange), and are occasionally later resolved as being either true (green), or false (red). Each line represents a rumour story, while the X axis represents the timeline.
Fig 6.
Distribution of delays (in hours) in resolving false (red) and true (green) rumours. Horizontal lines represent 25, 50, and 75 percentiles.
Fig 7.
Retweet networks representing retweets of unverified source tweets (orange), accurate tweets that support true rumours or deny false rumours (blue) and inaccurate tweets that deny true rumours or support false rumours (brown).
Table 3.
Percentages of retweets for unverified, accurate, and inaccurate tweets.
Fig 8.
Retweet timelines showing the percentage of retweets that each type of tweet gets in 15 minute steps.
Higher retweet percentages at the beginning represent a high interest in spreading the tweet in the very first minutes.
Fig 9.
Z-score values for retweets of different types of tweets, representing the extent to which different types of tweets receive retweets below or above the overall average of retweets that all rumours get.
Fig 10.
Distribution of support ratios.
Distribution of support ratios before and after resolving tweets for true and false rumours, as well as for rumours that remain unverified. Horizontal lines represent 25, 50, and 75 percentiles.
Fig 11.
Support heatmap representing the number of rumours that spark (1) more support, (2) more denials, (3) the same number of each, or (4) neither.
Fig 12.
Analysis of discussion around rumours.
Z-score values for the amount of discussion generated by different rumours, representing the extent to which different rumours spark discussion below or above the overall average. Red corresponds to false rumours, green to true rumours and yellow to unverified rumours.
Fig 13.
Analysis of support ratio on rumours.
Z-score values for support ratios observed in different rumours, representing the extent to which different rumours get support below or above the overall average.
Fig 14.
Distribution of certainty ratios.
Distribution of certainty ratios before and after resolving tweets for true and false rumours, as well as for rumours that remain unverified. Horizontal lines represent 25, 50, and 75 percentiles.
Fig 15.
Distribution of evidentiality ratios.
Distribution of evidentiality ratios before and after resolving tweets for true and false rumours, as well as for rumours that remain unverified. Horizontal lines represent 25, 50, and 75 percentiles.
Fig 16.
Analysis of support, certainty and evidentiality by follow ratio.
Average follow ratios for users who express different types of support, certainty and evidentiality, along with error bars that represent the 95% confidence interval.