‘We won’t have the sexiest AI, but everything it says is true,’ says Narrative Science

The hyperbole surrounding artificial intelligence is so thick these days, that it can be hard to break through the hype when talking with a vendor of product — one of the reasons reporting on AI is so bad, generally speaking. 

When one does take the time, however, to talk to a vendor, some conversations can be enlightening about the state of the art of the field, and about the trade-offs between scientific research and practical engineering. 

A recent article in The New Yorker might have given the impression that Chicago-based Narrative Science, a nine-year-old company, is a purveyor of today’s deep learning forms of artificial intelligence. 

But what the company really is, is a software firm that combines basic analysis of a domain of knowledge, such as business, with natural language tools refined over many years, to make programs that help people get stuff done. 

“Artful engineering, I like that term,” says Nate Nichols, distinguished principal of product strategy and architecture at Narrative, in an interview by phone with ZDNet.

“We won’t have the sexiest AI in the world, but everything it says is true, and everything is clear, and that’s what actually is valuable,” says Nichols.

Also: No, this AI can’t finish your sentence

What Narrative makes are software programs, two of them so far, Quill and Lexio, that take database information and turn it into natural-language sentences. It can handle query and answer pairs of the sort, “Who made the most sales last week? Suzie was the best, but Bob jumped from the bottom of the pack.” Quill has been around for years, but Lexio just came on the market in the past few months. 

This is not deep learning. “We are in Chicago, we have a real midwestern approach,” says Nichols. “We are not trying to compete with Google.”

The point of both Quill and Lexio is to move beyond the traditional dashboard of analytics software, which shows business information in pie charts and bar graphs, to instead deliver whole sentences that give an employee what they need to know at that moment.

“The salesperson or a customer success person, or someone on a shop floor — that person may have no idea what information they need, or what is relevant, or what is statistically significant,” explains chief executive Stuart Frankel. 

“I actually think that the head of sales is much more likely to absorb and consume and act on information if it’s delivered in a simple way versus a dashboard you give them,” says Frankel. A  lawyer by training, Frankel spent years as president of a marketing agency called Performics, later integrated into ad giant Publicis Groupe. He perhaps understands the dynamics of how humans relate to information better than some who have spent years building neural networks. 

Nichols, who earned a PhD in computer science from Northwestern University, and who has been with Narrative since the beginning, explains the multiple components that go into making what is “true and clear.”

It starts with what’s called an ontology, a collection of entities in a given domain, such as sales, and their relationships. For the time being, all that is kept in a SQL database, but in the future, putting it into a graph database such as that sold by Neo4J is a possibility he is “really excited about,” says Nichols. 

The ontology provides structure to everything that comes after it. “The system knows that pipelines exist” of widgets in factories, or sales prospects, say, or whatever is relevant, “and how things are entering and leaving” the pipeline, he explains. “And then, at runtime, what you want to say is all based on data.”

Then comes analysis of the data — automated, usually through very basic statistical tools, nothing particularly fancy, things such as linear regression. 

Also: OpenAI has an inane text bot, and I still have a writing job

“If a salesperson says, this deal is going to close by the end of November, we do a regression analysis, and see the deal is not going to close until late in December, it’s going to slip, based on the size of the deal versus deals like this that have taken X-long to close,” explains Nichols. 

The results of analysis get put into a “compositional natural language generation engine” that can build sentences and paragraphs, with rules about punctuation and other aspects of composition. 

“A lot of it is thinking pretty deeply about how stories are structured,” says Nichols, when asked where the technical challenges are. Some of that involves doing “part of speech tagging,” to see the nouns in phrases, to look for grammatical relationships between entities, to infer the actual semantic relationships, he explains.

But a lot more is the continual review of what’s produced by the programs, to refine how language is represented and how language is then generated for human consumption. There’s a constant feedback loop, “tens of thousands” of interactions by clients with the program. “As more and more of these kinds of database interactions happen — people interact with stories — we are learning a ton of what resonates.” 

“There are so many things that could be included in the story,” says Nichols. Something as simple as the expression “sales were up a big increase over the prior quarter,” he notes, is “really hard to figure out where that’s coming from semantically.” A slang characterization such as the salesperson “crushed it” this quarter, meaning, achieved positive results to a larger-than-normal extent, is the kind of characterization that is “super important for readers,” says Nichols. 

That sense of what’s important in language cannot yet be achieved by a deep learning model such as the “GPT2” neural network from OpenAI, says Nichols.

“GPT2 and such, that’s not something we do,” he says. “We are not Word2Vec,” he says, referring to the program for language “embedding” developed at Google in 2013. “We don’t predict the next letter in a sentence, that’s not something we will ever do, that’s not the kind of engine we are building.”

To anyone who might insist that Narrative is not what they conceive of as AI, that doesn’t bother Nichols. What is in that product is “the intelligence of the people who built the system,” he points out.

“For example, the system these days is a lot better at saying sales were such and such compared to the previous quarter, all that got really tighter,” he observes, “and that was driven by an engineer who saw that our time frame expressions were hard to read before and saw how to improve that.”

Maybe that sounds like a religious debate, but to Nichols, it’s a practical concern, namely: What is verifiable? Because Quill and Lexio come out of an ontology, rather than a probability distribution of language modeled in vectors, the sentences are connected to the truth of the database more transparently, he says.

“People reading blogs about GPT2 could be frightened at that running on top of their data, because those things have no basis in fact,” he says. “People are not excited about producing stories that are not related to reality, whereas our approach is very comforting and exciting.”

Chief executive Frankel reinforces the notion of transparency, based on his numerous conversations with clients. “We took this position when we started commercializing the technology,” he recalls. “We went into financial services firms, and when you try to sell to them, if it’s a black box, and you ask them to trust you, they say, No, we have to know how these apps are working.” 

That adds up to market share, in Narrative’s case, because it means the software can spread to more and more people in an enterprise who are not data scientists. Whereas Quill was integrated with analysis tools, such as Tableau and Qlik, Lexio will go deeper into tools that line-of-business people use, such as Salesforce. “The idea is that you go from data buried in Salesforce to the information you need, without this middle process of exploration and analysis,” he explains. 

Also: Going deeper: Moveworks snags $75M in new money to bolster deep learning in enterprise

Tools like Tableau, and others from vendors such as Alteryx, have long sought to “democratize” data science, by making it easier to do. That’s fine, says Frankel, but it only goes so far. “This notion of democratization of data is a fool’s errand,” says Frankel. “It’s a meritocracy; if you have access to tools and skills you can get a lot out of the data, so it’s less of a democracy than a meritocracy.” 

True democracy, in the Narrative view, is to get the info to that salesperson who, as he says, doesn’t even know what to look for or what analysis to perform.

“If you’ve got 100,000 people in an organization, only 25,000 are using BI tools,” he insists, using the acronym for “business intelligence,” the analytics dashboards. The remaining three-quarters of employees is a big opportunity for Narrative. In dollar terms, out of a total market for tools of perhaps $200 billion, Tableau and Qlik and Alteryx only cater to $50 billion.

To Nichols, it comes down to pride in making something that works on a practical basis, even if it isn’t fancy deep learning.

“My mom was a public school English teacher,” he recalls. “I remember how much time she spent grading tests and struggling with the school dashboard to see which students were keeping up.”

“If something like Lexio would help teachers, would let them focus on teaching, but be informed by the data, that’s a world that I want, those are all things I really believe in, empowering people with data.”