BACKGROUND IMAGE: iSTOCK/GETTY IMAGES
They say a picture is worth a thousand words, but sometimes, a little extra text can make that picture even more valuable. Enter natural language generation software.
"If you show someone a data point, and it's interesting, they should then ask why it is that way," said Luke Horgan, director of digital channel analytics at financial services company USAA. "NLG [natural language generation] answers that. It adds context to everything."
Natural language generation software is a subcategory of natural language processing. Advances in deep learning have opened up new abilities in machines to work with free text.
At first, the focus was on using algorithms to interpret text, but they're increasingly being used to create text themselves.
Natural language generation software beefs up reports
USAA is using NLG to automatically generate descriptions of metrics surfaced in its BI reporting tool. Horgan and his team are responsible for tracking the performance of digital products, such as the opening of new accounts online. USAA uses Adobe Analytics to track and report this data.
Then, software from Narrative Science looks at the data included in graphs and develops a one-paragraph description of why the numbers look the way they do. If the number of new accounts opened in a given day is lower than planned, the software can identify trends that might explain why.
Horgan said the dashboard itself has been around for a while, and it was fairly successful because it answered business users' primary questions about how their products, such as checking accounts and insurance policies, were selling. But that information then created more questions, which led to USAA's need for natural language generation software.
"The idea is to publish a dashboard that answers questions, [as well as] the questions that it generates," Horgan said.
Natural language generation tools start from scratch
But NLG isn't only about answering direct questions. It is also being used to create more in-depth pieces of text.
For example, at real estate listing site Trulia, natural language generation software automatically creates descriptions of neighborhoods for users evaluating properties.
Trulia's vice president of engineering, Deep Varma, said the project started with the marketing team writing its own descriptions of certain cities. He described this manual process as laborious, and said it wouldn't have scaled to all the locations where Trulia has listings.
So the company developed an NLG algorithm to automatically write these neighborhood descriptions. The algorithm that produces the descriptions is a neural network written in Python code. It pulls in data from publicly available data sources, as well as Trulia's own curated databases, and contextualizes information about a city's climate, population demographics and home prices.
Varma said, when his team started the project, there was a lot of discussion about whether the text created by machines would be good enough to pass for something written by a human. Today, the platform is working so well that few people in their internal testing can tell the difference. "When we started this project, AI was in an infancy stage," Varma said. "Now, it's delivering value."
See how healthcare experts are using NLP in care settings
Call centers still behind on NLP implementation
Natural language generation steps up its game