Where is NLG Most Successful Commercially?

Commercial NLG is growing rapidly, which is great.  I am sometimes asked where commercial NLG is most successful.  Ie, there are lots of great applications of NLG, but which sector is overall most important in commercial NLG?

Looking just at publically available information (I cannot say anything confidential about Arria), I believe that the most important sector for commercial NLG in late 2018 is in financial reporting.   This is perhaps because (a) there is a lot of money in finance, and (b) data and use cases are similar enough to allow systems to be replicated.

What are NLG Companies Focusing On?

All major NLG companies are involved in projects in a wide range of sectors and niches.  But what are they focusing on?

Arria: It is difficult for me to say anything concrete about Arria without violating commercial confidentiality.   So I will simply point out that Arria’s two latest press releases (at the time I am writing this) are about its work with Ernst and Young (one of the “Big Four” auditing and accountacy firms) and the release of version 2 of its NLG add-on for business intelligence tools.

Automated Insights: The Automated Insights home page lists “customer stories” in the areas of marketing analytics, sales analytics, earning reports, sports articles, and fantasy football.

Ax Semantics: The Ax Semantics home page focuses on product descriptions, for e-commerce and tourism.  It also mentions financial reporting and automatic journalism as case studies.

Narrative Science:  I note that most of the customers listed on Narrative Science’s home page are in the financial services area.  The NS home page also features a prominent blurb for their NLG for business intelligence tool.

Yseop: The Yseop home page stresses two types of NLG application, customer relationship management (CRM) and financial reporting.

Overall, it is striking that every one of the above companies is highlighting its involvement in financial reporting.  This is not true for any other sector.

Not Enough Money: Weather Forecasting and Journalism

What about weather forecasting, which is where commercial NLG got started?  CoGenTex launched its landmark FoG weather-report generator (Goldberg et al 1994)  (the first-ever deployed operational NLG system) back in 1992!

When we started Data2Text back in 2009, one of our first contracts was in weather forecast generation (some of this was published in Sripada et al 2014).  I mentioned this to someone who worked at CoGenTex in the 1990s, and this person warned me that “there is no money in weather forecasts”.   In other words, meteorological services is a very small slice of the global economy (miniscule compared to financial services), and also its dominated by government agencies which tend to build systems in-house and/or buy them from local companies.  So although NLG can do impressive and useful things in weather forecasts, its perhaps not the ideal sector for making large amounts of money.

I suspect the same is true for automatic journalism (robojournalism).  There are a few niches, such as financial and sports reporting, where there is money and indeed demand for NLG.   But overall the journalism industry is in deep trouble from a financial perspective, as advertising and “eyeballs” move online, so there is not a lot of money to spend on NLG technology.   Similar to weather forecasts, I think NLG can do impressive things in journalism which help society (and Arria is involved in automatic journalism projects), but I suspect it is not the best sector for making significant profits.

Not Enough Consistency: Medicine

What about using NLG in medicine, to support clinicians or patients?  This is something I am very keen on; I think NLG can make a real contribution to both informing clinicians (to enhance decision making and reduce medical errors) and to informing patients (to support decision making and lifestyle changes, and to reduce stress).  And the healthcare sector of course is huge, so there is plenty of money to chase.

Two major problems, though, are inconsistent datasets and processes, and liability if something goes wrong.  Both of these came up when I talked to a medical informatics company many years ago about commercialising our Babytalk research  (this was a university project, which pre-dated Arria).  We had developed an NLG system which worked in the Neonatal Intensive Care Unit (NICU) in the main Edinburgh hospital, and generated (completely automatically) reports for doctors, nurses, and parents.   So anyways, the first question the medical informatics company asked us was “How hard would it be to deploy Babytalk in another hospital?”.  And the answer was that it would be a lot of work to deploy Babytalk even in a nearby hospital such as Glasgow, because of major differences in equipment, data collected, and clinical processes.   And even more work to deploy Babytalk in USA (the biggest healthcare market), because of the different healthcare system (eg, in USA billing/insurance issues become very important).   This clearly was not what the company wanted to hear.  They wanted software which could be sold to lots of hospitals, not software which required major modifications for each hospital it was deployed in.

And I think this problem is very common in healthcare.  There is little standardisation in clinical practice, datasets, IT systems, etc, and this makes it challenging to widely deploy advanced IT systems, including NLG.   Actually the best place to find consistency may be in consumer-facing health apps and gadgets; this is currently a tiny part of the overall healthcare market, but it is growing very rapidly.

The other issue that came up in our discussions about Babytalk, and in medical NLG (and AI) more generally, is liability.    If something goes wrong and a patient suffers (or even dies) because of what the IT system said, who is liable?   An enormous headache and “can of worms”, for anyone in this sector.   In general, society is far less tolerant of AI systems making mistakes than of human clinicians making mistakes.

Conclusion

NLG is most commercially successful when it is deployed in sectors which (A) are large, and (B) have consistent data sets and use cases.  At the moment financial reporting seems to fit these requirements best. But of course this is a generic statement, there are plenty of exciting use cases for NLG in other sectors!

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s