Contextualising trends

Chennai: Twitter used contextualising trends as a key tool during the recently concluded state elections to fight misinformation on the platform, a senior executive said.

The results of polls in five states – Uttar Pradesh, Goa, Manipur, Punjab, and Uttarakhand – will be out on March 10.

“Contextualising trends was a key tool for us during the elections in Japan this year (and) for the current regional elections in India, too,” said Leonardo Stamillo, director, editorial curation at the US-based microblogging platform said on Tuesday in a virtual interaction with the media. “Our creators are now partnering with machine learning engineers to improve recommendation models, and have even more trends contextualised for our users.”

He said Twitter started creating content to contextualise trends in a bid to give users context about what was trending.

This ‘experiment’ began two years ago before the US presidential elections and the company has been seeing really great results so far, Stamillo said.

Curation can provide additional context on trending topics by writing headlines, descriptions and also pinning a tweet that represents the conversation happening in that particular trend, he said.

“According to an internal survey, in the US, 85% of the respondents said that the trend description helped them understand why a trend was trending. And 82% said descriptions make using trends on Twitter a better experience,” he added.

Twitter will continue to invest in the languages it detects misinformation in, said Yoel Roth, head of site integrity at Twitter.

Twitter evaluates content for violations in languages including Arabic, English, French, Hindi, Japanese, Korean, Portuguese, Spanish and Turkish, he said, adding that it was working to expand to more languages soon.

“Twitter is also testing new ways to empower people who use Twitter to let the company know when they’ve come across misleading information,” Roth said. “In August, Twitter began testing a new misinformation reporting feature in the US, Australia, and South Korea, and in January expanded this test to Brazil, Spain and the Philippines.”

He said Twitter has learned that these reports are ‘helpful but noisy.’

On average, Roth said that about 10% of misinformation reports reviewed through this flow were actionable — compared to 20-30% for other policy areas.

There were also several “off-topic” reports that don’t contain misinformation at all, he said. For instance, several reports of “voting misinformation” were related to contests or votes about people’s favourite musical artists.

“Twitter will continue to experiment with this feature (rather than roll it out to everyone) because it needs more research and training data, especially outside of English to help optimize how Twitter uses these reports,” Roth said.

“Twitter’s already seeing clear benefits from reporting for identifying patterns and trends in misinformation — especially when it comes to non-text-based misinfo, such as media and URLs linking to off-platform misinformation.”