National / Media

Automation comes to news-gathering in Japan

by Tomoko Otake

Staff Writer

The news business in Japan has long been notoriously labor-intensive. Reporters assigned to the crime and disaster beats have largely relied on briefings from police officers or firefighters for their initial reports.

But to get a scoop or avoid falling behind their competitors, they have also toiled away at youchi asagake (visiting officials at their homes late at night or in the early morning for brief solitary chats with them).

In an era where word of a fire or accident quickly spreads through social media, however, the value of such practices, at least for routine news stories, has been thrown into question. In recent years, major media outlets have instead turned to a “news agency” that doesn’t even have one journalist. Instead it is staffed by engineers who develop technologies to let machines be their news hounds.

JX Press Corp., headed by 29-year-old entrepreneur Katsuhiro Yoneshige, offers live feeds of witness accounts of crimes, accidents and other emergencies posted on social media to all five major TV networks in Tokyo, NHK and Kyodo News, as well as several regional newspapers.

The firm’s AI-powered computers filter through a deluge of tweets, Facebook posts and Instagram shots around the clock, pick up potential news events and pinpoint what is happening where. They also send alerts about potential breaking news to clients.

“What’s hot right now is a fire in the city of Ise, where seven or eight houses are apparently burning and an evacuation order has been issued,” Yoneshige said in late January at his office in Chiyoda Ward, Tokyo, casually scanning the firm’s news alert software, Fastalert, on his laptop. “Oh, and this is unconfirmed information, but some amount of money seems to have been withdrawn from a cryptocurrency exchange called Coincheck.”

Later that night at a hastily arranged news conference in Tokyo, Coincheck announced its computer system had been hacked and ¥58 billion worth of its customers’ assets had been stolen, sending shock waves through the nation’s nascent but rapidly expanding cryptocurrency market.

Yoneshige, who set up his company in 2008 while still in college in Tokyo, said he had long been interested in the news business, creating a niche website on airline-related news when he was still in high school.

But it was through watching the boom and bust of online journalism platforms in Japan in the 2000s, such as OhmyNews, JanJan and PJNews, that really made him question the existing business models for news organizations.

Such news sites initially offered high hopes for citizen journalism in the digital age, but they closed down in a few years, due mostly to financial reasons.

“I realized how hard it is for online media to monetize themselves,” he said. “And I wondered if there was any way I could solve the cost structure issues of the news media.”

His solution: Make the news-gathering process more efficient and cheaper by letting computers take some of the burden.

The cost of introducing the firm’s Fastalert software, which was launched in 2015, is about the same as hiring one rookie reporter, he said.

“Considering the cost and time involved in having reporters call the police every hour, monitor radio communications or search for tips on Twitter, we have managed to cut the labor significantly.”

Though Yoneshige himself is not an engineer, 17 of his 24 employees are. (The remaining staff are in sales and administration.)

The core of the firm’s technology lies in judging the type of news and its location by analyzing text, images and video posted by social media users. For instance, even if a Twitter user doesn’t say where a fire is taking place, the location can be instantly guessed through their images, such as a city name painted on a passing firetruck or a street sign in the background, he said.

In the early days, the system was prone to error. For example, tweets by people at yakiniku (Korean barbecue) restaurants were often mistaken as reports about fires. Such glitches have been fixed by feeding in “teacher data” to filter out possible non-news information, Yoneshige said.

He is quick to emphasize that his job is not to drive journalists out of a job. Though Fastalert offers tips, it’s ultimately the job of the media to confirm the information.

As with Western news outlets, where automated robots have crept into newsrooms and have been churning out sports and business earnings stories, a day will soon arrive in Japan when bots will master the art of writing simple straight news stories off press releases, he said. But in-depth analyses will still be the work of humans, and such content will remain in huge demand, he stressed.

For example, companies with favorable financial results tend to go into detail in their earnings reports, boasting why they did so well, he said. But when the results are bad, they often start out with a general statement about the state of the economy and hide behind vague language.

“When robots summarize such reports, they make little sense,” Yoneshige said. “That’s where reporters need to dig further, touching base with sources or delving into the background behind the reports. That’s what only humans can do and what machines have no prospect of achieving, at least in the near future.”