While the robots are coming for journalists, might natural language generation also make inroads into financial, legal and medical analysis?
Software-written journalism attracts a lot of coverage – possibly a little more than is justified, by journalists fearful of their jobs. Nevertheless, last year saw the Los Angeles Times run articles about small earthquakes generated by its Quakebot algorithm, while the Associated Press now produces more than 1,000 company results stories a month using software.
But suppliers of natural language generation (NLG) systems are keen to apply their technology to any kind of writing based on structured data. While the opportunity/spectre of automatic journalism has provided good publicity for such software, it is financial and business services, healthcare and intelligence that are driving its growth – although those in the field say its full potential has yet to be discovered.
“We are still exploring. We’re still discovering where the technology is best used,” says Ehud Reiter, professor of computing science at the University of Aberdeen and chief scientist for supplier Arria. “My own opinion is that where NLG is really good is where you have information from a lot of sources that needs to be combined, integrated and presented to various audiences.”
It is particularly useful in making raw data accessible, he adds. “There are a lot of places where people want the data wrapped up into a story. That’s where NLG is really powerful.”
Narrative Science started as a university project called StatsMonkey, which automatically wrote reports on baseball games based on statistics. “We have got a tremendous amount of press around the potential disruption to journalism from software like this,” says Nick Beil, chief operating officer of Narrative Science. “The reality is that it has not been a focus of our business for over three years.”
The firm decided to focus on financial and business services, and its Quill software has customers including Deloitte, Credit Suisse and the US Automobile Association. It also works with US intelligence agencies through an investment from In-Q-Tel, the CIA’s venture capital firm.
In August, Narrative Science launched a version of Quill to generate portfolio reviews, so clients receive a narrative rather than, or in addition to, a set of graphs and tables. When written by humans, such commentary is likely to affordable only for wealthier clients, but automating it opens it up. “It really helps them scale their service,” says Kim Neuwirth, product manager for financial services.
Personalise online interactions
One financial services customer, USAA, is using Quill to personalise online interactions in the way its staff can do on the telephone. “Member calls are still part of our core operating model, but we’re working to understand how to create personalised experience and build relationships in those growing digital situations,” says Zack Gipson, USAA’s chief innovation officer.
“One of the ways to do that is to take the same kind of information we would get from a member call and do that in a digital sense. Narrative Science can take that data and create context and relevance.”
USAA is also looking at using Quill internally for reporting and analytics, to improve productivity by allowing staff to spend more time on other work. Gipson says the firm is exploring other potential uses and will use the software “wherever it makes sense”.
Another use is in satisfying increasing demands for regulatory returns. Deloitte uses Quill to produce the narrative section of its suspicious activity reports for US regulator FinCen, an anti-money laundering return introduced in 2012 that must be completed within 60 days.
Deloitte says in a white paper that large financial institutions can generate thousands of these returns annually. Narrative Science has just launched Quill for Anti-Money Laundering, which automates processes in this area.
Providing financial advice
But Beil says clients would be hesitant about having Quill providing financial advice. “Quill is being used for the diagnostic and descriptive – here’s what happened and why it happened,” he says. “We’ll leave the financial advice to the adviser, or the robo-adviser,” the latter being software specifically designed to provide advice.
But there is potential for NLG to provide advice in less regulated areas, says Beil.
Automated Insights has followed a similar route to Narrative Science, having gained publicity by producing automatic sports reports that it distributed through hundreds of websites, thousands of Twitter accounts and Facebook pages and mobile apps. “It gave us the opportunity to showcase that we could both sound human and produce at scale,” says chief revenue officer Adam Smith. “One of the exciting things with sports is that you can experiment with tone.”
The firm’s Wordsmith software produces Associated Press’ automated journalism and is used for financial and business writing, but also generates descriptions of property for sale and descriptions of products for e-commerce. The firm is now encouraging more experimentation through a beta programme.
Devices for elderly people
One user is US firm GreatCall, which sells simple mobile devices for elderly people who want to continue to live independently. The devices are packaged with operator and medical advice lines and a GPS tracking service that enables emergency help to be sent.
Chris Gutierrez, director of product innovation at GreatCall, said many devices are bought by those caring for elderly people, often their children. “They are shockingly busy,” he says of the carers, with heavy work and family commitments, sometimes living some distance from the user.
“What they needed was a quick snapshot of how the parent was doing, that they had the device, had it on, if they were going out and taking care of themselves. They weren’t going to look at pages of analytics.”
So the firm developed GreatCall Link, which summarises data from the service using NLG in just three sentences. “Tone was absolutely critical,” says Gutierrez, with care taken to respect the client, such as saying “the device wasn’t charged” and that the carer may want to offer a gentle reminder about this.
The summary provides a brief description of where the user has been, based on the GPS data and identifications of favourite locations, although this can be turned off. Gutierrez says some elderly users have been tracked visiting new girlfriends or casinos.
Carers can receive the summaries online or through an app. “It’s certainly one of our most engaging features,” says Gutierrez. “The ability to analyse data sets, either visually or in charts, is not universally shared.”
Summarising such data, and trends within it, in words “has tremendous value”, he adds, and GreatCall is developing new care-givers services using Wordsmith which it plans to launch in 2016.
GreatCall does not say in the summaries that they are written by software, says Gutierrez. Doing so would undermine their brevity and he doesn’t believe the readers would be concerned. Other users, including Associated Press, disclose the use of Wordsmith.
Either way, there are risks in trying too hard to sound human, says Automated Insights’ Smith. “In some instances, people want to add movie quotes, or specific words that stand out,” he says, but the firm advises against this.
“Canned text may originally have been written by a human author and appear warm and fluent,” says Paul Piwek, a senior lecturer in the Open University’s faculty of maths, computing and technology. “But when the machine puts together individual words, the result will not always be as fluent as that of a native human speaker.”
NLG systems are taking over work that may previously have been done by junior staff as part of their training. But Narrative Science’s Neuwirth, who previously worked as a financial analyst, says such software can also be used to help train staff.
“You actually have an automated way to store and save the thought process behind how your best analysts work, and let other analysts gain a viewpoint into that,” she says.
Those involved in NLG tend to see it as a way to use human staff more effectively, rather than replacing them. Associated Press uses Wordsmith to increase its coverage of smaller companies, but also to allow human journalists to write more complex stories. “They are able to provide colour, quotes and context,” says Smith. “The machine can do the what and when; the human can do the why and how.”
People in the loop
The University of Aberdeen’s Reiter adds: “I personally feel it’s better to use NLG with people in the loop. There are certain things people are really good at and certain things computers are really good at.”
Humans can synthesise ideas at a high level, while computers can do routine work very quickly and keep a full record of what data lies behind each phrase, he says. “Ideally, you want to combine the two.”
Some kinds of writing may be technically suitable for NLG, but may annoy people. The University of Aberdeen used NLG to give adults with poor skills feedback on tests, and “a few people got really upset when the computer told them they had problems adding”, says Reiter.
That also applies to sensitive news, he adds: “Some kinds of information should come from a person looking you straight in the eye.”
Although healthcare involves many such sensitive messages, it also has potential for NLG. From 2011, the University of Aberdeen worked with the NHS Lothian health board on using software to turn data from an electronic patient record system into reports at the Royal Infirmary of Edinburgh’s neonatal intensive care unit.
There were three kinds of report: a live summary for doctors and nurses to enable decision-making, a handover document for nurses at changes of shift, and information for parents. The work ended when the board changed its patient record software.
Biomedicine and life sciences
The Open University’s Piwek says biomedicine and life sciences could use NLG to make data accessible for quality control of the data by human domain experts. There is also academic research into text summarisation, paraphrasing and simplification – all part of a field more broadly known as natural language processing – with potential for wider use.
But some areas are likely to resist automation, specifically those without adequate machine-readable information, says Piwek. “Text types that are likely to be problematic include those based on evidence and information from a range of sources in different formats, such as documents, interviews and experiments, and which require interpretation, human judgement and creativity to combine this information, and find it in the first place,” he adds.
These would include “a good piece of investigative journalism, a scientific paper that explores new ideas and a recommendation document from a consultant on business strategy”, says Piwek.
Writing these through software would either require human experts to express their insights in data format – which could take longer than in words – or will have to wait until the dawn of full artificial intelligence, which, says Piwek, is still at least three decades into the future.
So, although software-generated writing based entirely on structured data is with us, it seems likely that many of the readers, the interviewees and the writer of this article are safe from having their work taken over by machines – at least for now.
First published by ComputerWeekly.com, 27 November 2015