Increasingly, companies are turning to artificial intelligence to understand what people say about their products and services on Twitter or Facebook or elsewhere. The goal is to react more quickly to complainers and perhaps sell more stuff to happy customers.
But with sarcasm, there is a big gap between what people say and what they mean. And, because computers tend to take everything literally, they simply don’t get the joke.
For example, “You look wonderful,” can mean two very different things depending on the context and the speaker. It could mean that you do, in fact, look great, or it could mean the opposite coming from the late comedian Don Rickles, with his trademark exaggerated eye roll.
Get Data Sheet, Fortune’s technology newsletter.
This is a tough problem for AI to handle, but one that conversational analytics startup Gong, based in Tel Aviv with U.S. offices in Palo Alto, Calif., hopes to solve.
Companies wanting to fine-tune online recommendations, for example, would like to know if a user who wrote, “I loved this movie” is serious or flippant. It’s a problem that film information site The Internet Movie Database tries to overcome using an algorithm that analyzes people’s comments to automatically create a rating of one to 10.
But, a user who says a movie is “fantastic” could generate an excellent review that really does not reflect the user’s intent if he or she meant it sarcastically, Lotem Peled, chief data scientist for Gong, tells Fortune.
The problem is that “input may be sarcastic but output is not sarcastic” she adds. She therefore created a neural network, basically a system that collects conversational data and automatically tries to make sense of it, without programmers having to intervene much.
Any sort of AI project needs to test its theories and hypothesis, so Peled built her own data set from a massive number of tweets over the past few months. During the lead up to the U.S. presidential election, there was “lots of sarcasm out there” to work with, she joked.
Peled is interested in researching problems where AI technology still lags human intelligence. “AI is very advanced in some aspects, but in terms of understanding human nuances it’s still pretty far off.” She conducted research at Technion Israel Institute of Technology, and now Gong is incorporating it in its own software.
By getting a better handle on what people really mean, Gong thinks it can help do some very pragmatic things like helping sales people close more deals. To accomplish that, Gong records phone or online conversations between sales people and customers or would-be customers.
By analyzing information about who speaks longer—the buyer or seller—as well as pauses in the conversation, along with how many deals get done, Gong says it can help sales teams sell better. Gong technology can thus be used in conjunction with business software from Salesforce (crm) and other companies, she noted, that teams use to track sales and target potential customers.
“It’s not quantitative but qualitative data,” said Peled, whose project, which she calls the “Sarcasm Sentimental Interpretation GeNerator” (SIGN), is detailed here.
Basically, Peled’s system looks at conversations and parses not only words for what is positive or negative as well as pauses between speakers and words. then it maps the conversations to sales outcomes. Did this chat lead to an additional sale or not?
To be fair, virtually every software company, including Salesforce (crm), Microsoft (msft), Google (googl)and others are talking a lot about integrating AI smarts into their products. Gong’s secret sauce thus far appears to be its ability to suss out sarcasm
There are also important noncommercial uses for this research. Improving AI’s understanding of sarcasm could help people with Autism or Asperger’s syndrome, who often have difficulty discerning whether others are being sarcastic, by cuing them in to the meaning of what people say to them.
There are other nuances that AI could address better as well. For example, knowing when words are meant literally or figuratively can avoid a lot of problems.
Anyone who remembers the late, great Get Smart sitcom of the 1960s may recall that Hymie the Robot was overly literal. When someone told Hymie to “kill the lights,” for example, Hymie would shoot out the lights.