Sure, there are many ways to describe how one team won or lost the game, but the list of descriptors is still finite. There are only so many ways to say the 49ers sucked this year.
In 2013, Carl Frey and Michael Osborne published a paper titled “The Future of Work,” which considered this question: How susceptible are jobs to computerization? Using an algorithm based on “things which are hard to computerize,” namely social intelligence (PR), creativity (fashion design), and perception and manipulation (surgery), they estimated the likelihood of computerization for 702 distinct jobs taken from the U.S. Bureau of Labor Statistics. The outcome: 47 percent of U.S. jobs fell into the “high risk of automation” category, with more than a 70 percent chance of being automated over the next decade or two. (You can read the paper for yourself at tinyurl.com/oj67kae.)
The primary driver of this new wave of automation? Computers, via machine learning, have become adept at both pattern recognition (for example, identifying faces in photographs) and manipulation of objects (for example, training robotic arms to pick up small objects through trial and error). With these new skills, computers/robots are more likely to take over jobs which Frey and Osborne describe as “non-routine cognitive tasks,” such as driving vehicles and writing news articles.
Wait a second. Writing news articles?! That’s awfully close to writing columns‑like this one. But when you think about it, certain kinds of news‑sports, for example‑follow fairly identifiable formats. Sure, there are many ways to describe how one team won or lost the game, but the list of descriptors is still finite. There are only so many ways to say the 49ers sucked this year.
Automated Insights (www.automatedinsights.com) sells a program called “Wordsmith,” which will automatically create text from data, using templates. Every quarter, the Associated Press (AP) uses Wordsmith to create 4,400 stories on corporate earnings from raw earnings data. Using manual techniques, the AP was only able to provide reporting on a few hundred companies. On the sports side, the AP recaps all Triple-A, Double-A and Class A games‑about 10,000 games per year‑using Wordsmith as well.
Automated Insights is not alone in the Natural Language Generation (NLG) market. Narrative Science (www.narrativescience.com) is a Chicago-based competitor. They sell “Quill,” software designed to “create data-driven communications at machine scale.” In other words, in a world of “big data,” programs such as Wordsmith and Quill attempt to identify patterns in data from a specific domain (e.g. finance), and generate a written description of those patterns in natural language customized for a particular audience. This BBC news story (written by a human, we suppose) provides a closer look at examples of how this works, using stories written by both Wordsmith and Quill: tinyurl.com/j3lqugk.
Think you can tell the difference between a story generated by a human and one generated by software? The New York Times offers this quiz to test your perception: tinyurl.com/kb6t3qn. Even knowing a bit about the subject, I only identified five of eight examples correctly, scoring higher that 52 percent of those taking the quiz. Hmmph.
But, you say, even though the people who write news stories about financial results and sports may lose their jobs, there will be new jobs created for people who customize Quill and Wordsmith for particular domains and audiences. After all, ATMs led to a rise in the number of bank tellers employed, didn’t it? Alas, the people who lost their jobs writing articles are unlikely to be the same people customizing apps, so individually there’s a lot of hardship. And though there may be more bank tellers, it’s not clear whether they are making as much money as they used to (and demand is expected to decline in the future). Assuming we re-train those forsaken article-writers to be app-customizers, then supply and demand means that app-customizer wages will go down. It’s certainly unclear that automation will bring benefits to all, except those who work for the companies doing the automation itself (they have jobs), and the owners of the companies that automate their operations (they see lower costs).
Editor’s Note: This month’s Tech Talk column was written by “Mossberg,” a machine-learning program developed under the direction of Dr. Ramsden J. Farnor, head of the machine learning program at UCSF. Mossberg, named for the famed Wall Street Journal tech columnist Walt Mossberg, generates columns based on current technology news data. The magazine would like to thank Dr. Farnor and UCSF for providing access to this cutting-edge technology, which promises to save NorthBay biz a significant portion of what it presently pays columnists.
