patpitchaya - Fotolia
Natural language generation, a field in artificial intelligence which automatically turns facts and statistics into coherent English and other languages, offers important advantages for finance and accounting, according to Robert Dale, a longtime expert in the technology.
By submitting your personal information, you agree that TechTarget and its partners may contact you regarding relevant content, products and special offers.
Since 2012, Dale has been CTO and chief strategy scientist at Arria NLG, based in London, and one of the first companies to offer automatic text generation. He is also co-author, with Arria's chief scientist Ehud Reiter, of the textbook, Building Natural Language Generation Systems, first published in 2000 and then again in 2006.
In an interview with SearchFinancialApplications, Dale discussed NLG, its benefits for the office of the CFO and the future of artificial intelligence (AI).
Natural language generation software is best used anywhere with large sets of data. What distinguishes your technology?
Robert Dale: We use a technology that we call precision NLG. What that is really all about is basically building sentences very carefully, taking into account their grammatical structures. We do that programmatically from data. This is significant if you have data that varies significantly from one instance to another. For example, if you are reporting on two parties with wealth portfolios that are drastically different, then it becomes very difficult to use a template to provide those narratives. With the kind of technology we have developed in NLG, you get creative control.
How can natural language generation software be used in the office of the CFO?
Dale: It provides the ability for the CFO to automate reporting for other parties. The CFO is typically very familiar and happy with numbers and tables and charts. But the information a CFO has to send to other people in an organization is often best provided via language. It is quite an onerous task for a human being to take large quantities of numerical data and turn that into something that is easy to consume for a non-specialist. There is a real role for the technology in any situation where the CFO needs to communicate to a wider audience.
Can it be used for compliance or would NLG create a compliance issue?
Dale: NLG can help with compliance. There's lot of compliance that requires narrative reporting. In terms of whether NLG can do the right thing and comply with the rules, so to speak, it is very straightforward to build an NLG system that can explain the reasons it undertook for reaching its conclusions. It can actually aid compliance.
Arria NLG is offering a new reporting tool. What kind of reporting can NLG provide?
Dale: We just released a software as a service (SaaS) product called Recount. This is aimed primarily at small- to medium-sized businesses. It hooks up to existing financial analysis and accounting packages and it provides to users a narrative explanation, or a health check, about how their business is doing. It looks at a number of key performance indicators (KPIs) around cash flow, sales and accounts receivables, for example, and relays that information on the numbers back to the user in narrative.
Should the use of natural language generation software be disclosed to a customer in customer communications or in articles?
Dale: If we don't do that, I think eventually people will insist that we do. It is likely that there will be a demand from regulators that reports should be basically signed by the application that has written them. We are not in that situation now, but I can see why people might want that to be the case.
Why would regulators and customers want that?
Dale: People like to know the provenance of the information that is being delivered. For example, when you apply for a mortgage or a loan, there is often an automated assessment process that decides whether you should have that loan. European Union law now requires that the reasoning of the assessment be explained. This is a direct reaction to the use of lots of machine learning technology in these spaces. People want to know why they did not get a loan. These same sorts of forces are going to demand that the kind of text that we generate is signed by the machine that generates it so people know they are being delivered information not by a human, but by a machine.
Can natural language generation software allow for personalization? Can it provide individualized customer communications?
Dale: Mass personalization is one of the key benefits of the technology. In terms of personalization, most of the variations are data driven. If I know something about your demographics, your bank account or your stock portfolio, then the information I deliver will be tailored to your particular needs based on those data variables.
What do you see as the shortcomings of NLG? What needs to be improved?
Dale: At the moment, we still have to tell our machines quite a lot about how to use language. We have not worked out all the rules. No one has. We can build pretty sophisticated systems that can produce pretty sophisticated narratives, but we still cannot do that with quite the nuance of a human author. A human author has intuitions about how to use language which are currently beyond our capabilities. We are getting there step by step, but there is still a bit of a distance to go. At this stage, you are not going to see a computer write better than the bestselling novelist, for example. That's still a few years away.
Natural language generation software is a field in artificial intelligence. How far away are we from peak artificial intelligence?
Dale: You are going to get different answers from different people. I think we are a long way from peak AI. I think the kind of AI systems we have built so far are still very much what I would call 'silo systems.' They work in vary narrow spaces. You can have an AI system that can chat to you about insurance or you can have an AI system that can drive a car, for example, but they are all different systems. Today's AI systems are very specialized with quite specific tasks. We are a very long way away from general intelligence, which is how I would think of the notion of peak AI. Once we get to general artificial intelligence, then I think we will have made that target. I would define peak AI basically as having built a machine intelligence that is completely open-ended in terms of the skills it can acquire and use.
Natural language generation software makes headway in finance
Artificial intelligence machines on verge of wide adoption
Consider the true situation of artificial intelligence technology