Facebook’s flood of languages leave it struggling to monitor content
NAIROBI/SAN FRANCISCO (Reuters) – Facebook Inc’s struggles with hate speech and other types of problematic content are being hampered by the company’s inability to keep up with a flood of new languages as mobile phones bring social media to every corner of the globe.
The company offers its 2.3 billion users features such as menus and prompts in 111 different languages, deemed to be officially supported. Reuters has found another 31 widely spoken languages on Facebook that do not have official support.
Detailed rules known as “community standards,” which bar users from posting offensive material including hate speech and celebrations of violence, were translated in only 41 languages out of the 111 supported as of early March, Reuters found.
Facebook’s 15,000-strong content moderation workforce speaks about 50 tongues, though the company said it hires professional translators when needed. Automated tools for identifying hate speech work in about 30.
The language deficit complicates Facebook’s battle to rein in harmful content and the damage it can cause, including to the company itself. Countries including Australia, Singapore and the UK are now threatening harsh new regulations, punishable by steep fines or jail time for executives, if it fails to promptly remove objectionable posts.
The community standards are updated monthly and run to about 9,400 words in English.
Monika Bickert, the Facebook vice president in charge of the standards, has previously told Reuters that they were “a heavy lift to translate into all those different languages.”
A Facebook spokeswoman said this week the rules are translated case by case depending on whether a language has a critical mass of usage and whether Facebook is a primary information source for speakers. The spokeswoman said there was no specific number for critical mass.
She said among priorities for translations are Khmer, the official language in Cambodia, and Sinhala, the dominant language in Sri Lanka, where the government blocked Facebook this week to stem rumors about devastating Easter Sunday bombings.
A Reuters report found last year that hate speech on Facebook that helped foster ethnic cleansing in Myanmar went unchecked in part because the company was slow to add moderation tools and staff for the local language.
Facebook says it now offers the rules in Burmese and has more than 100 speakers of the language among its workforce.
The spokeswoman said Facebook’s efforts to protect people from harmful content had “a level of language investment that surpasses most any technology company.
But human rights officials say Facebook is in jeopardy of a repeat of the Myanmar problems in other strife-torn nations where its language capabilities have not kept up with the impact of social media.
“These are supposed to be the rules of the road and both customers and regulators should insist social media platforms make the rules known and effectively police them,” said Phil Robertson, deputy director of Human Rights Watch’s Asia Division. “Failure to do so opens the door to serious abuses.”
ABUSE IN FIJIAN
Mohammed Saneem, the supervisor of elections in Fiji, said he felt the impact of the language gap during elections in the South Pacific nation in November last year. Racist comments proliferated on Facebook in Fijian, which the social network does not support. Saneem said he dedicated a staffer to emailing posts and translations to a Facebook employee in Singapore to seek removals.
Facebook said it did not request translations, and it gave Reuters a post-election letter from Saneem praising its “timely and effective assistance.”
Saneem told Reuters that he valued the help but had expected pro-active measures from Facebook.
“If they are allowing users to post in their language, there should be guidelines available in the same language,” he said.
Similar issues abound in African nations such as Ethiopia, where deadly ethnic clashes among a population of 107 million have been accompanied by ugly Facebook content. Much of it is in Amharic, a language supported by Facebook. But Amharic users looking up rules get them in English.
At least 652 million people worldwide speak languages supported by Facebook but where rules are not translated, according to data from language encyclopedia Ethnologue. Another 230 million or more speak one of the 31 languages that do not have official support.
Facebook uses automated software as a key defense against prohibited content. Developed using a type of artificial intelligence known as machine learning, these tools identify hate speech in about 30 languages and “terrorist propaganda” in 19, the company said.
Machine learning requires massive volumes of data to train computers, and a scarcity of text in other languages presents a challenge in rapidly growing the tools, Guy Rosen, the Facebook vice president who oversees automated policy enforcement, has told Reuters.
GROWTH REGIONS
Beyond the automation and a few official fact-checkers, Facebook relies on users to report problematic content. That creates a major issue where community standards are not understood or even known to exist.
Ebele Okobi, Facebook’s director of public policy for Africa, told Reuters in March that the continent had the world’s lowest rates of user reporting.
“A lot of people don’t even know that there are community standards,” Okobi said.
Facebook has bought radio advertisements in Nigeria and worked with local organizations to change that, she said. It also has held talks with African education officials to introduce social media etiquette into the curriculum, she said.
Simultaneously, Facebook is partnering with wireless carriers and other groups to expand internet access in countries including Uganda and the Democratic Republic of Congo where it has yet to officially support widely-used languages such as Luganda and Kituba. Asked this week about the expansions without language support, Facebook declined to comment.
The company announced in February it would soon have its first 100 sub-Saharan Africa-based content moderators at an outsourcing facility in Nairobi. They will join existing teams in reviewing content in Somali, Oromo and other languages.
But the community standards are not translated into Somali or Oromo. Posts in Somali from last year celebrating the al-Shabaab militant group remained on Facebook for months despite a ban on glorifying organizations or acts that Facebook designates as terrorist.
“Disbelievers and apostates, die with your anger,” read one post seen by Reuters this month that praised the killing of a Sufi cleric.
After Reuters inquired about the post, Facebook said it took down the author’s account because it violated policies.
ABILITY TO DERAIL
Posts in Amharic reviewed by Reuters this month attacked the Oromo and Tigray ethnic populations in vicious terms that clearly violated Facebook’s ban on discussing ethnic groups using “violent or dehumanizing speech, statements of inferiority, or calls for exclusion.”
Facebook removed the two posts Reuters inquired about. The company added that it had erred in allowing one of them, from December 2017, to remain online following an earlier user report.
For officials such as Saneem in Fiji, Facebook’s efforts to improve content moderation and language support are painfully slow. Saneem said he warned Facebook months in advance of the election in the archipelago of 900,000 people. Most of them use Facebook, with half writing in English and half in Fijian, he estimated.
“Social media has the ability to completely derail an election,” Saneem said.
Other social media companies face the same problem to varying degrees.
Facebook-owned Instagram said its 1,179-word community guidelines are in 30 out of 51 languages offered to users. WhatsApp, owned by Facebook as well, has terms in nine of 58 supported languages, Reuters found.
Alphabet Inc’s YouTube presents community guidelines in 40 of 80 available languages, Reuters found. Twitter Inc’s rules are in 37 of 47 supported languages, and Snap Inc’s in 13 out of 21.
“A lot of misinformation gets spread around and the problem with the content publishers is the reluctance to deal with it,” Saneem said. “They do owe a duty of care. “
This post originally appeared on Reuters.com
