1
Illustration by Joe Magee
Philana Patterson, assistant business editor for the Associated Press, has been covering business since the mid-1990s. Before joining the AP, she worked as a business reporter for both local newspapers and Dow Jones Newswires and as a producer at Bloomberg. “I’ve written thousands of earnings stories, and I’ve edited even more,” she says. “I’m very familiar with earnings.” Patterson manages more than a dozen staffers on the business news desk, and her expertise landed her on an AP stylebook committee that sets the guidelines for AP’s earnings stories. So last year, when the AP needed someone to train its newest newsroom member on how to write an earnings story, Patterson was an obvious choice.

The trainee wasn’t a fresh-faced j-school graduate, responsible for covering a dozen companies a quarter, however. It was a piece of software called Wordsmith, and by the end of its first year on the job, it would write more stories than Patterson had in her entire career. Patterson’s job was to get it up to speed.

Patterson’s task is becoming increasingly common in newsrooms. Journalists at ProPublica, Forbes, The New York Times, Oregon Public Broadcasting, Yahoo, and others are using algorithms to help them tell stories about business and sports as well as education, inequality, public safety, and more. For most organizations, automating parts of reporting and publishing efforts is a way to both reduce reporters’ workloads and to take advantage of new data resources. In the process, automation is raising new questions about what it means to encode news judgment in algorithms, how to customize stories to target specific audiences without making ethical missteps, and how to communicate these new efforts to audiences.

Automation is also opening up new opportunities for journalists to do what they do best: tell stories that matter. With new tools for discovering and understanding massive amounts of information, journalists and publishers alike are finding new ways to identify and report important, very human tales embedded in big data.

Algorithms and Automation

Years of experience, industry standards, and the AP’s own stylebook all help Patterson and her business desk colleagues know how to tell an earnings story. But how does a computer know? It needs sets of rules, known as algorithms, to help it.

An algorithm is designed to accomplish a particular task. Google’s search algorithm orders your page of results. Facebook’s News Feed determines which posts you see, and a navigation algorithm determines how you’ll get to the beach. Wordsmith’s algorithms write stories.

AP’s Philana Patterson helps train software to write earnings stories

AP’s Philana Patterson helps train software to write earnings stories

In order to write a story, Wordsmith needs both data about the specific task and guiding principles about the general one. Your GPS needs to know where you are now and where you’re going; it also needs to know that “giving directions” means showing the fastest route from point A to point B, which depends on a variety of other data like whether streets are one way, what the speed limits are, and if there’s traffic or construction. Similarly, to write an earnings story, Wordsmith needs the specific data about a company’s quarterly earnings, and it also needs to know how to tell an earnings story and what information it needs to accomplish that goal.

To train Wordsmith, Patterson had to think about the possible stories the data might tell and which metrics might be important. Did a company report a profit or a loss? Did it meet, beat, or miss analyst expectations? Did it do better or worse than it did in the previous quarter or a year earlier? Deciding which metrics and data might matter was a head-spinning task. “You have to think of as many variables as you can, and even then you might not think of every variable,” she says.

Working with other journalists on the business desk, she settled on a handful of storylines, with all their accompanying variety. She then worked with software developers at Automated Insights, the Durham, North Carolina–based company behind Wordsmith, who translated those story models into code the computer could run to create a unique story for each new earnings release. Today, the AP produces about 3,500 stories per quarter using the automated system, and that number is set to grow to more than 4,500 by the year’s end. Automation is taking off, in large part because of the growing volume of data available to newsrooms, including data about the areas they cover and the audiences they serve.

The race to publish business data quickly dates back to the first Lloyd’s List in 1734

The history of the news business is, in some ways, a history of data. The ability to collect and publish business-critical information faster than others has been a key value proposition since Lloyd’s List was first published in London in 1734. Companies like Bloomberg and Thomson-Reuters have built empires on their ability to provide market data to business readers. But even outside the business media landscape, data has been an important part of why customers have turned to news outlets: Box scores, weather, election results, birth and death announcements, and poll results are all classic elements of a newspaper.

Just as media have undergone a digital revolution, so have the data that inform many elements of the news. Information of all types is increasingly accessible in the form of “structured data”—predictably organized information, like a spreadsheet, database, or filled-out form. This makes it well suited for analysis and presentation using computers.

The growth of structured data is at the heart of increasing automation efforts. Business and sports have long been data-intensive coverage areas, so it’s no surprise that automation is being used in these areas first. The sports and business agate were among the first items to exit the print pages and find new homes online, in part because this kind of information is easily handled by digital systems. But today, a growing volume of private and public data is available in digital formats, and new tools make it easier to pull data out of even non-digital formats.

But data isn’t the same as information. Algorithmic content creation isn’t just about turning a spreadsheet of numbers into a string of descriptive sentences; it’s about summarizing that data for a particular purpose.

How a Robot Learns to Write a Story

An annotated Associated Press earnings reports by Automated Insights’ Wordsmith system

“Our goal prior to automation was to have 130 words onto the wire within 15–20 minutes of the press release” announcing a company’s earnings for the most recent quarter, says Philana Patterson, the AP’s assistant business editor. Since it began using Automated Insights’ Wordsmith platform in 2014 to automatically generate these stories, the AP has been able to get stories of up to 500 words onto the wire as quickly as one minute after the earnings data is released. The Wordsmith software is capable of generating up to 2,000 stories a second, says Joe Procopio, chief product officer at Automated Insights. Click on the highlighted text below to see how it works:

FedEx reports fourth-quarter loss of $895 million, [1]falls short of [2]forecasts

1. Phrases like “falls short” are part of a strict vocabulary the Associated Press defined for Wordsmith based on the same style guide used by AP reporters. Patterson says the AP’s goal was to let the numbers tell the story and only use plain verbs. The choice of word is partially randomized, with “beats,” “misses,” and “matches” having a higher probability of being used than “tops,” “falls short of,” and “meets.”

2. The AP compares performance to Wall Street forecasts as well as to the company’s own announced expectations, the latter of which is typically considered a more powerful indication of performance.

MEMPHIS, Tenn. (AP) _ FedEx Corp. (FDX) on Wednesday reported [3] fiscal fourth-quarter loss of $895 million, after reporting a profit in the same period a year earlier.

3. Wordsmith analyzes numerous factors to find a story’s lede. There are hundreds of possible data points that could be used, based on comparisons among factors like current earnings data, the company’s historical data, performance of similar companies, or Wall Street expectations, Procopio says. This is similar to the guidelines given to the AP’s reporters for covering earnings: Net income (also called “profit” and “earnings”) is considered the primary benchmark for a company’s success, so it should be compared with the same quarter in the year prior to get a sense of whether the company is doing better or worse.

On a per-share basis, the Memphis, Tennessee-based company said it had a loss of $3.16. Earnings, adjusted for non-recurring costs and asset impairment costs, were $2.66 per share.

The results did not meet Wall Street expectations. The average estimate of 12 analysts surveyed by [4]Zacks Investment Research was for earnings of $2.70 per share.

4. Wordsmith relies on data entered by humans at Zacks Investment Research for AP stories, but reporters will often incorporate additional quotes and context when covering the same earnings report. AP reporter David Koenig wrote a story on this release that included explanations for why FedEx’s revenue was down: lower fuel surcharges, a strong dollar, and a change in accounting of pension costs. Patterson says the AP hasn’t been able to get those contextual details into the structured format Wordsmith needs quickly enough, so the Automated Insights stories are produced without them. The AP always assigns additional reporting resources to 80 companies, while another 220 are reviewed each quarter by editors. “We’re trying to use earnings stories as a window into the company’s strategy,” AP business editor Lisa Gibbs says of the work her reporters are doing now. “When you’re having to churn out hundreds of earnings stories you don’t necessarily have the time or the brainpower to look for those more interesting tales.”

The [5]package delivery company posted revenue of $12.11 billion in the period, [6]which also fell short of Street forecasts. Six analysts surveyed by Zacks expected $12.39 billion.

5. After using “FedEx Corp.” on first reference, the company is referred to as a “Memphis, Tennessee-based company” or a “package delivery company.” To avoid repetition, Wordsmith pulls those descriptions from a database of company locations and business types that have been defined in accordance with AP style.

6. Since Wordsmith has already given FedEx’s earnings and noted that it did not meet expectations in the previous paragraph, here it notes that overall revenue missed expectations as an aside using the word “also,” since that detail is of secondary importance to data about earnings. It is able to do this because separate algorithms check the overall flow of the language at the sentence, paragraph, and story levels to keep it from becoming repetitive. In effect, the system is writing new sentences based on what it has already written, Procopio says, similar to how a human might write.

For the year, the company reported [7]profit of $1.05 billion, or $3.65 per share. Revenue was reported as $47.45 billion.

7. Because Wordsmith is monitoring the flow of the story, this can lead to small changes in the language, such as substituting synonyms for words that occur multiple times. Terms like “profit” and “earnings,” for instance, may be used interchangeably in some situations. “We don’t ever want an article that’s ‘profit, profit, profit, profit,’” Procopio says.

[8]FedEx expects full-year earnings in the range of $10.60 to $11.10 per share.

8. This data about FedEx’s expected earnings for the year is not always included when Zacks sends its initial data. When additional data from Zacks is delivered, the entire dataset will be re-evaluated and a new story will be generated.

FedEx shares have risen roughly 5 percent since the beginning of the year, while the Standard & Poor’s 500 index has increased almost 2 percent. The stock has climbed 31 percent in the last 12 months.

[9]This story was generated automatically by Automated Insights (http://automatedinsights.com/ap) using data from Zacks Investment Research. Full Zacks research report: FDX.

9. Every automated AP story goes out on the wire with this disclaimer as well as a link to a page that gives more detail about how the automation works. If a reporter has rewritten an automated story, the disclosure will remain, with the language updated to read: “Elements of this story were generated partly by Automated Insights.”

The Associated Press’s data is provided by Zacks Investment Research; that company uses human analysts who review Securities and Exchange Commission data, stock pricing, and press releases to build a custom feed of the numbers the AP has requested. That data is sent to Automated Insights, and Wordsmith assembles the stories following the rules Patterson and her colleagues helped set.

Translating even the simplest data means converting the loose guidelines a human reporter might follow into concrete rules a computer can follow. For example, a human reporter might have a general idea of when a company’s performance was very different from analyst expectations, based on their knowledge of the industry. But for the algorithm, the AP had to specify exact ranges for which the spread between actual earnings and expectations is considered large or small. Wordsmith uses such metrics to decide both which words are used to describe the data and how the story is structured— for example, whether the fact that a company missed analyst estimates should be mentioned in the headline. The story-assembling algorithm uses a predetermined set of vocabulary and phrases (known as a corpus) that follows the AP’s strict stylebook rules.

“It’s a lot!” Patterson says. “To come up with a system to trigger the right type of story, we as reporters and editors and programmers have to figure out this stuff ahead of time.”

You have to know, “what it is you want your data to tell you,” says Evan Kodra, a senior data scientist with Lux Research, a Boston-based market research firm. The more targeted and specific the questions, the better the results. “It still takes a lot of creativity to define the problem.”

Editors say that’s one reason they’re incorporating automation technologies into their workflow: It enables them to focus on the fundamental work of being a reporter. “Isn’t that our whole job: understanding the purpose of any kind of narrative before we do it?” asks Scott Klein, an assistant managing editor at ProPublica. “In a way, our job is figuring out the purpose of the story and figuring out a way of telling it.”

ProPublica’s Scott Klein says automation gives reporters more time to figure out the best way to tell a story

ProPublica’s Scott Klein says automation gives reporters more time to figure out the best way to tell a story

ProPublica’s first—and so far only—foray into automated journalism was part of “The Opportunity Gap,” a data-driven analysis of which states are (or aren’t) providing low-income high school students with the coursework they need to attend and succeed in college.

Studies have shown that advanced high school coursework can improve a student’s college outcomes, and in 2011, ProPublica released an investigation into where low-income high school students have equal access to, and enrollment in, advanced courses. The analysis was based on a new data set from the U.S. Department of Education. ProPublica used the data to create an interactive news app to accompany the story. Website visitors could explore the data at the federal, state, district, and school levels.

Two years later, the team was preparing to update the app with current data when Narrative Science, a Chicago-based competitor of Automated Insights, approached them. The company’s platform, Quill, uses a similar algorithmic method to produce stories from sets of data. ProPublica had spent months analyzing, interpreting, reporting out, and correcting the Department of Education’s data set. “The data were so well structured and we understood it so well,” says Klein, meaning it was a good fit for automation. They decided to use automation tools to provide a written narrative to accompany each of the 52,000 schools in the database.

How an Algorithm Creates a Story

Each of the profiles needed to provide a summary of the data for an individual school, but it also needed to connect each school with the broader story. To provide context, ProPublica decided to include both a summary paragraph outlining the thesis of the broader investigative work and a comparison with another school to show the local context. To produce the narratives, ProPublica’s editors provided Narrative Science with their complete data set as well as some sample write-ups. But the most important part was selecting the right schools for comparison.

The editors wanted to prioritize comparisons that showed differences with regard to opportunities, but it wasn’t appropriate to compare a school in California to a school in Chicago because the economic and policy conditions can vary widely across such geographic gaps. Based on their reporting, Klein and data editor Jennifer LaFleur decided to first restrict the comparison to schools within the same district or state, before highlighting data that showed similarities or differences between the compared schools. “Even though the data look the same,” says Kris Hammond, chief scientist and co-founder of Narrative Science, “there are so many different environmental conditions that are outside the scope of this data that the comparisons would not fly and would, in fact, be making false analogies.” This kind of journalistic insight is critical to finetuning the performance of algorithms.

Like any human reporter, robot journalists need editors. But the challenge of editing automatically generated stories isn’t in correcting individual stories; it’s in retraining the robot to avoid making the same mistake.

In May 2015, The New York Times wrote an article about a new study on how where you grow up affects your economic opportunities later in life. The study used tax records to track the fates of 5 million children who moved among U.S. counties between 1996 and 2012. The study concluded, “The area in which a child grows up has significant causal effects on her prospects for upward mobility.” To accompany the article, the Upshot team produced an interactive piece that highlights data for each of the 2,478 counties included in the study.

But rather than just present a searchable database or zoomable map, graphics editors wrote an article that adapts to the user, based on their current location. By looking at the user’s IP address, key paragraphs highlight local income statistics and compare them to national averages. The accompanying map is automatically focused on that county and its neighbors. Users can choose other locations, but rather than seeing an entirely separate story, the same story gets new data and a new lede for the new location.

When “The Best and Worst Places to Grow Up” was released, many users didn’t notice that the text was assembled algorithmically. They just arrived at the page and thought their version of the story was the only version of the story. That seamless experience is partially the point, but it comes with its own editorial demands. “Because people think this is edited by a human editor, you have to have the same standards, accuracy, quality, and tone. There’s a big danger in messing things up,” says Gregor Aisch, a graphics editor for The New York Times.

For its best places to grow up report, The New York Times used algorithms to assemble stories that varied depending on the user’s location

The story uses pre-assigned blocks of text and follows specific rules for how to assemble paragraphs based on the available data. In some cases, it might be as simple as substituting a new number or county name. For the best- and worst-performing counties, the story got additional sentences that only appeared in those contexts. In addition to editing the pre-written chunks of text, editors had to check for flow between sentences in multiple possible arrangements.

The challenge is even trickier for newsrooms using systems like Quill or Wordsmith, because these systems use more “word variables,” and they have more options for how to describe data. So, the same data might be able to produce a dozen different variations of a story.

For now, the process for editing these stories is more or less the same as for human writers: reviewing drafts. Klein says most of the drafts that ProPublica received at first had errors. Data appearing in wrong parts of the story was the most common mistake. Once editors mark up the drafts, developers make the changes to the code to ensure it doesn’t happen again. Over time, ProPublica felt confident it had a system that produced accurate stories and used language with which its editors were comfortable.

The Associated Press also spent months reviewing drafts, refining the story algorithms, and verifying the quality of the data supplied by Zacks. The first quarter that the system was live, editors reviewed drafts of every story before it was put out onto the wire, checking for errors in both the data and the story. Now, the majority of stories go live on the wire without a human editor’s review.

The AP says the only errors it still sees come from errors in the data passed to the system. Some are simple typos or transposed numbers, while others depend on more complicated human errors. Unless data is gathered by a digital sensor, the process almost always starts with humans doing data entry, which is often where problems are introduced. Since the project’s inception, Patterson says, only two published errors have been traced back to the algorithm.

In July, Netflix released second-quarter earnings at the same time as its stock underwent a 7-to-1 split. But the data Wordsmith received didn’t reflect the split, so Wordsmith initially reported that the price of an individual share fell 71 percent and noted that the company had missed analyst expectations for per-share earnings. Neither of which was true. In fact, investors who owned the stock saw an increase in the value of their portfolio; Netflix’s share price has more than doubled since the beginning of the year. This was, in effect, a human error: The analyst data should have reflected the stock split. But Wordsmith does not have an automated warning that kicks in when something anomalous—like a 71 percent drop in share price from a company like Netflix—appears. The lesson, for automated and human-generated stories alike: Your data have to be bulletproof, and you need some form of editorial monitoring to catch outliers.

The story was updated with a correction, following the same processes as for any human-generated story. But because of the way AP stories are syndicated, uncorrected versions of the story persist online. Patterson says it’s wrong to blame automation for that kind of error. “If the data’s bad you get a bad story,” she says.

Tom Kent, the AP’s standards editor, acknowledges that mistakes are an issue that the AP takes seriously—but he also points out that human-written stories aren’t error free, either. “The very stressful job for a human of putting together figures and keeping data sets separate and not mixing revenue and income and doing the calculations correctly was a prescription for mistakes as well,” he points out. According to Patterson, who oversees all corrections (human or otherwise) for the business desk, the error rate is lower than it was before automation, though she declined to provide exact figures.

Algorithm-assisted Journalism

The imperative to avoid errors prompted the AP to keep its automated stories simple, which makes them, well, somewhat lifeless. Other Wordsmith users include descriptions of major factors, such as most point-contributing players for fantasy sports or top-performing stocks and categories for financial portfolio summaries, that influence the overall trend in a data set. The AP has opted to exclude these more analytical facts, often included in human stories, from automation because of concerns about adding too much complexity too quickly. “There are things we decided not to do quite yet that were presented as possibilities,” Patterson acknowledges. “We chose not to add them to the stories, because we were really committed to making sure that the accuracy of the stories was intact.”

Instead, the Associated Press has human editors who add context to many of its automated stories. At least 300 companies are still watched closely by the AP’s business desk staff. There are 80 companies that always get additional reporting and context by Associated Press staff; another 220 get reviewed by editors, who may enhance the story with their own reporting or context. That system has created significant efficiencies for the AP, freeing up 20 percent of the staff’s time across the business desk, estimates Lou Ferrara, the vice president and managing editor who oversaw the project. And that doesn’t take into account the impact of the initiative on the AP’s customers.

At AP, automated stories have freed up 20 percent of the business desk’s time

One of the biggest impacts of the Associated Press’s automated earnings project has been its expanded coverage of smaller companies that are primarily of interest to local markets. The AP’s customers are, largely, local outlets, and companies of interest to these clients had fallen out of AP coverage during the cutbacks of the 2000s. For communities, this was a potentially significant loss. “If there’s a big company, it’s employing people in your family, your neighbors, people you go to church with,” says Patterson. “There are a lot of people who are interested in the economic health of that company.”

In Battle Creek, Michigan, for example, the Kellogg Company is one of the region’s most important employers, and its fingerprints are all over town—from thousands of monthly pay stubs at the bank to the names on a school to the W.K. Kellogg Foundation’s cheery-looking headquarters downtown. Pat Van Horn is among the locals who worked at Kellogg until she retired in 2010. She and her husband Lance still pay attention to what’s happening at the company. They have friends who work there, and like many locals, they’ve got Kellogg stock in their portfolio.

So each quarter, when the company releases its earnings statements, the Van Horns glance at the Battle Creek Enquirer to see how things are going. “You know, what was the earnings report this quarter, the dividends are going to be X amount per share,” says Lance. “We follow them a little bit.”

The Enquirer uses the Associated Press’s earnings stories as the foundation for its coverage of the company; local reporters add context, digging deeper on issues that are likely to impact Battle Creek. That frees up reporters and editors to do the work that the computers can’t do.

For the AP, content licensing is king, making up the vast majority of the company’s revenue, and newspaper and online customers accounted for 34 percent of 2014 revenue. Continuing to deliver content those customers want is key to retaining their business. “It’s not like we’re going to be growing revenue in the local markets in any particular way,” says Ferrara. Instead, the AP sees the automated earnings as a way to retain customers, particularly those that have been hard hit by job losses across the industry.

The AP is doubling down on that strategy; the company has continued to expand its automation efforts, adding public companies with a market capitalization above $75 million as well as select Canadian and European firms. Many of these companies would never have been covered by the AP’s staff writers. The same is true for other areas of coverage the AP is looking to automate, including Division II and Division III college football and basketball games.

By offloading the basic reporting work, the AP hopes it’s making it easier for local papers to focus on the stories that matter to community members, like the Van Horns. “We’re not here merely to be just churning out numbers,” says Lisa Gibbs, the AP’s business desk editor. “We’re really writing these stories for customers who are more likely to have shopped at a Walmart than to own individual stock in Walmart.”

Gibbs was hired just after the introduction of the automated earnings stories. She says it was an opportunity for the team to rethink how the company was going to cover business. With automation, Gibbs says her team has been able to focus on doing the kinds of medium-sized enterprise stories that had been squeezed out before. She points to the example of a piece by business reporter Matthew Perrone, who covers the U.S. Food and Drug Administration, which reported on a lack of regulatory oversight for the growing number of stem cell clinics. “We were able to take some time, send him to travel to some of these clinics, and ultimately publish the story,” she says.

Alexis Lloyd of  The New York Times R&D Lab says the public’s thinking about automation hasn’t changed since the 1950s

Alexis Lloyd of The New York Times R&D Lab says the public’s thinking about automation hasn’t changed since the 1950s

The AP isn’t alone in using automation to support its reporting efforts. One of the first places to adopt automation in the newsroom was the Los Angeles Times. The paper’s Homicide Report maintains a database with information about every homicide reported by the Los Angeles County coroner’s office; each victim profile includes a brief automatically generated write-up. It’s up to reporters to decide which stories deserve more in-depth reporting.

The LA Times built on the lessons from that project with the introduction in 2011 of Quakebot. Ken Schwencke, a digital editor on the LA Times’s data desk at the time, used data from the USGS Earthquake Notification Service to automatically generate short reports on earthquakes above the “newsworthy” threshold of a 3.0 magnitude. LA Times reporters review the stories, publish them, and update the story with additional information as it becomes available. Quakebot’s big advantage is speed; a story can be posted online in under five minutes. This kind of assistive role is one that many news organizations insist is the foundation of their automation efforts.

The AP recently hired its first “news automation editor,” Justin Myers. He sits on the editorial team and is represented by the News Media Guild, just like his editorial colleagues. His job is to help figure out how to streamline editorial processes and “give time back to the writers, editors, and producers, who in a lot of cases are slogging through whatever processes we’ve built up over the years, rather than focusing on doing journalism.”

Myers has spent his first few months on the job mostly interviewing reporters, editors, and producers to find out what work he can help take off their plates. The number one question he’s asking: “How do you spend your time?” If it’s possible to automate some of a staffer’s burdensome tasks, Myers is happy to help. “Let’s have a computer do what a computer’s good at, and let’s have a human do what a human’s good at,” he says.

Alexis Lloyd, creative director of The New York Times R&D Lab, agrees. The general public’s thinking about automation hasn’t been updated since the 1950s, she says. Typically, we imagine an all-or-nothing scenario: all with humans or all with machines. She says that’s wrong; across all kinds of industries the approach to automation has changed to focus on more assistive technologies. “We’ve been thinking that the future of computational journalism and automation will—and should—be a collaborative one, where you have machines and people working together in a very conversational way,” she says.

Several news organizations are using automation to support their reporters’ work behind the scenes, too. Lloyd mentioned Editor, a new tool that integrates with the company’s content management system to help reporters tag content by providing automated suggestions. Similar efforts are under way at BBC News Labs, with a tool called Juicer.

“The future of computational journalism and automation will—and should—be a collaborative one.”
—ALEXIS LLOYD, CREATIVE DIRECTOR OF THE NEW YORK TIMES R&D LAB

These tools support news organizations in their push to develop new storytelling formats that highlight the relationships between news events and help provide readers with richer context. Most of these efforts require large amounts of detailed metadata that can help link together stories that have in common people, places, or ideas. Adding metadata is a frustrating task for most reporters, who are typically more concerned with crafting their story than dissecting it. Automation is a way to expand the use of metadata—without putting an extra burden on reporters and editors.

RELATED ARTICLE
Jonathan Stray on simple robot reporting tools

Behind-the-scenes tools can also help reporters in more proactive ways. For example, another tool from The New York Times R&D Lab automatically tracks its stories on Reddit, looking for hot conversations, and alerts journalists when there’s an active discussion of their work they might be interested in monitoring or participating in.

This is one of the most promising areas for automation in the newsroom, says Nick Diakopoulos, an assistant professor at the University of Maryland’s Philip Merrill College of Journalism, who has been studying the use of algorithms in media. By tracking social media or other public data sets, automation tools can help support newsgathering in a digital environment. Using automation tools like these can raise journalists’ awareness of issues, help them pay attention to important data sets, or listen to conversations and react more quickly, he says.

Personalization and Revenue

Automation can also become a useful tool for connecting with audiences more directly. In June, journalists at Oregon Public Broadcasting (OPB) rolled out a news app to accompany a series on earthquake preparedness in the state. The app, called Aftershock, provides a personalized report about the likely impacts of a 9.0 magnitude earthquake on any user’s location within the state, based on a combination of data sets.

The scenario isn’t just speculative; the region is widely expected to face a massive quake of just this sort, known as the Cascadia quake. “OPB has been doing a bunch of coverage on the Cascadia quake and how people can prepare, but a lot of people don’t care about a topic until it affects them directly,” says OPB’s Jason Bernert. “When you put them in the center of the story, they take an interest.”

Oregon Public Broadcasting’s automated localized earthquake preparedness reports woke up listeners to their vulnerabilities

Aftershock uses data sets on earthquake impacts that were modeled by the Oregon Department of Geology and Mineral Industries and impact zones defined by the Oregon Resilience Plan report. The data mixes and matches ratings for things like shaking, soil liquefaction, landslide risk, and tsunamis. In total, there are 384 possible combinations, and users see a version of the story that’s relevant to the location they’ve selected. As with The New York Times’s “Best and Worst Places to Grow Up” interactive, the news app dynamically stitches together the various elements of the story—which the OPB team calls “snuggets,” a portmanteau of “story nuggets”—based on the data for each location.

Some of the data applies to broad regions of the state, but other data sets have estimated impacts for regions as small as 500 meters. Aftershock takes advantage of that granularity by showing users the expected impacts for specific addresses. “There’s a big difference between ‘a 9.0 earthquake in Oregon’ and ‘your area is where shaking is going to be the worst,’” says Bernert. “It has a different emotional response for people to start different conversations.”

The editorial appeal of projects like this is clear—but personalization has the potential to attract the interest of the business side as well. Following the mid-July publication of The New Yorker’s in-depth article about the Cascadia quake, Bernert says, Aftershock’s traffic soared. For a few days afterward, the site was handling 300 times the usual number of requests. Other OPB reporting on the Cascadia quake saw an increase in traffic, but “the real social driver was Aftershock,” he says. On Facebook, users were sharing Aftershock and saying, “This is what’s going to happen to me; I better go out and get prepared” and encouraging others to check out how they would be impacted as well.

Although none of the current implementations have focused on monetization efforts specifically related to personalized content, it is the aspect of automation that could have the largest effect on potential news revenue.

Teach Your Computers Well

Avoiding the risks of algorithmic bias

Systems like Wordsmith and Quill are natural language generation (NLG) platforms. That means they’re designed to turn data into human-sounding prose. NLG is an active area of technology development, aimed at helping translate the growing stores of structured data into human-usable information.

Siri, Apple’s conversational assistant, uses natural language generation to answer simple questions and respond to user requests. But Siri is also using another technology that Quill and Wordsmith aren’t: natural language processing (NLP).

Natural language processing is a class of artificial intelligence tools that let computers analyze unstructured data—like newspaper stories—to identify patterns. Those patterns can then be applied to new situations. That’s how Siri figures out whether you’re looking for a restaurant or a pet shop when you ask her for the nearest “hot dog.” Then, Siri uses natural language generation to direct you to the nearest frank.

By looking at large volumes of data, NLP systems develop statistical models about how often humans use specific vocabulary, syntax, and phrases in a particular context. To do so, they depend on training data sets, known as corpora, that have something in common—e.g., all of the text is weather reports or earnings statements.

The models generated by NLP tools will replicate the same biases inherent in the corpora. Usually, that’s good: You want Siri to bias her response toward an edible treat, not a panting pet. But it has its dark side, too, and the risks of ignoring biases are significant. Google’s new Photos tool recently highlighted the issue when a black user found that photos of him and his friend were labeled “gorillas” by the software. The error was likely linked to a lack of black faces in the algorithm’s training data for “people.” And a number of recent studies have found issues with Google’s search algorithm: an image search for “CEO” returns almost exclusively images of men, while women conducting online job searches are shown fewer ads for high-level coaching services. These kinds of issues happen because the algorithms learned from the patterns of past user behavior—reinforcing damaging cultural stereotypes.

Today, Wordsmith and Quill require editors to explicitly construct the models that shape stories. But NLP-based systems could “read” multiple examples of the text editors want to create, alongside the data sources for the story, and develop their own statistical models for what makes an earnings story or a homicide report. Examples of this approach are few and far between outside of a research context—for now. But it’s not wise to bet against improvements in this area of technology. If news organizations adopt more artificial intelligence techniques, it will be important to ensure that they’re using diverse training data that reflect their efforts to produce more inclusive coverage of communities.

Automation is already being used today to personalize some news organizations’ homepages or to provide “recommended for you” features. By further increasing engagement with users, automation that personalizes content could have positive impacts on revenue from advertising and subscriptions. This kind of personalization provokes anxiety among many news professionals, who worry that personalization will limit readers’ exposure to the stories editors might deem important in favor of things that are frivolous. As Mark Zuckerberg said when describing the value of the News Feed, “A squirrel dying in front of your house may be more relevant to your interests right now than people dying in Africa.”

Automated Insights’ Joe Procopio, right, with company CEO Robbie Allen, says personalization is a growth area

Automated Insights’ Joe Procopio, right, with company CEO Robbie Allen, says personalization is a growth area

For now, most article personalization efforts focus on types of users, much as Aftershock uses automation to match specific addresses to general scenarios. It’s more like having a shirt with the right collar and sleeve-length measurements than a handmade, custom-tailored one. “We haven’t got it down to the person yet,” acknowledges Joe Procopio, chief product officer for Automated Insights.

That’s true of even the most sophisticated algorithms that are used by credit agencies, retailers, and personnel companies; vast quantities of personal info are crunched to pigeonhole users as a “type” that can be used to predict loan default risk, send perfect-for-you coupons, or characterize your management style. Take the example of Crystal, an artificial intelligence tool that helps you write better e-mails for specific individuals, based on their online profiles. The program reviews things someone has written online, such as their LinkedIn profile, and identifies them as one of 64 types. Each type has associated communication tips about things like vocabulary to avoid, how much detail to include, and how formal the language should be.

Automated Insights provides this kind of customization to its commercial customers already—one car-sales website uses Wordsmith to show users slightly different descriptions of the vehicles based on their profile, Procopio says. A first-time car buyer might be shown a description that emphasizes the car’s fuel performance, while a mother in the market for a family vehicle might see descriptions that emphasize safety ratings. In both cases, the information in the profiles is the same, but different features are prioritized.

Diakopoulos says lack of data is a significant barrier to such personalization of stories. To push further into true personalization, news organizations would need to collect a lot more information about their users—and develop strategies for how to address stories to those different types of users. “News organizations aren’t very good about even having user models,” Diakopoulos points out. “They don’t really know who’s on their site. That’s very different than having a robust user profile and an ability to adapt the page based on the cookie profile and so on.”

Even if users were to agree to provide more detailed information—by logging in with Facebook or LinkedIn, say—journalists would still need to be at the helm of efforts to target those users with content in specific ways. The complexity involved in automating a story with just one variable—geography—would become exponentially more difficult. For newsrooms, that presents significant challenges for editing, fact checking, and writing multiple variations of the “snuggets” to be used in the stories. “There is concern,” says Procopio, “that someone is going to read a story and not get all the facts because it’s biased toward that person. I don’t think that concern is warranted.”

As the technology improves, the potential value of personalization from a revenue perspective will certainly become more important. Frank Pasquale, a professor at the University of Maryland Francis King Carey School of Law and author of a recent book on the pervasive power of algorithms, “The Black Box Society,” argues that if stories can eventually be customized for users based on factors like their income, where they live, or any of the micro-categories (e.g., “cat lover,” “Walmart shopper,” or “STD sufferer”) that data brokers collect from our online lives, newsrooms will almost certainly face pressure to do so. “That’s going to be seen, eventually, as revenue maximizing,” Pasquale says.

He suggests a question for newsrooms to consider as they apply personalization: “To what extent is this ‘dead-squirrel’ personalization and to what extent is this personalization that draws people creatively into stories about other parts of the world?”

To focus on the latter, one option is to rely less on broad personal data that sparks fears about who algorithms assume a user is and instead focus on what’s relevant about a user’s relationship to a particular story. For example, The New York Times’s Upshot team has recently published a few stories that use in-story interactions to adapt a story to a user’s existing knowledge or views on the subject.

In one case, users were asked to draw a line on a graph they thought represented college enrollment rates across economic groups. Based on the line drawn, users were shown one of 16 different versions of the story, each of which explained the real data while comparing them to the user’s own assumptions about the issue. It was a simple but very successful piece of explanatory journalism because it focused the written article on information most relevant to the reader, without changing the reported parts of the stories. Projects like these also have the advantage of built-in transparency about what characteristics are being used to automate the story’s creation. Users actively provide the information to the system in order to get the information they want, and that input data is clearly linked to the story itself.

Small news organizations could, in fact, have the most to gain from automated journalism. They are well-positioned to draw on local data to write stories

Transparency is one of the stickiest issues facing automation systems, particularly as they intersect with personalization. Kent, the AP’s standards editor, thinks concerns about algorithmic transparency are overblown when it comes to automatically generating content. “Human journalism isn’t all that transparent,” he says. “News organizations do not accompany their articles with a whole description of what was on the journalist’s mind that could have affected his thinking process, whether he had a head cold, had just been hung up on by a customer service rep of the company he was writing about, and so on.”

Because the rules governing how automated stories get assembled are available for scrutiny, automated journalism may be more transparent than stories written by humans, he argues. But for the majority of projects, it’s hard to know what value readers might find in disclosures, even if they were presented. Mike Dewar, a data scientist in The New York Times R&D Lab, has written about the futility of publishing documentation if none of the intended audience can read it. Instead of publishing just open source data or documentation on algorithms, he argues, the community needs to adopt common standards and procedures.

That kind of standardization could benefit non-technical users, who would become more familiar with how such projects work and what to expect. Standardization could also help smaller newsrooms experiment with automation. At OPB, Aftershock was a hugely successful project, but it required some heavy lifting from the small public media team. Bernert and his colleague Anthony Schick built the app during a three-day build-a-thon sponsored by the University of Oregon’s journalism school, with the pro-bono assistance of a local interactive design firm, students, and academics. “There’s a lot of value to this kind of work,” Bernert says. “But how do we make it sustainable for a small public media newsroom?” Having a larger shared set of the technologies and methodologies would help.

Small news organizations could, in fact, have the most to gain from using automation. While Wordsmith and Quill are focused on expanding in big-dollar markets like financial information and insurance, they’ve demonstrated their technology on a variety of local data, such as water quality reports from public beaches and public bike-share station activity. Local news organizations could be well positioned to take advantage of this kind of structured data using automation, either by expanding their coverage or by creating new products. Commercial providers once siphoned off some of news organizations’ most important revenue streams by finding better ways to deliver classified ads, job listings, home sales, and other information—much of which is available in the form of structured data. Automation could be one way for news organizations to recapture some of that revenue.

After all, automation is about putting narratives around data, and news organizations have the skills and experience needed to do just that.

Show comments / Leave a comment