Chatbots and virtual assistants are getting a lot of attention in the AI world these days, and for good reason, as some of these tools are developing impressive skills. But there's more to the technology than conversational ability, and enterprises are finding value implementing component pieces of virtual assistants -- automated bots and natural language processing -- in other ways.
"If it's hard for Siri to figure out [natural language chat], it's going to be hard for us," said Brian Canty, CEO and co-founder of Lea, a San Francisco-based company that has developed a Facebook messenger bot. Lea -- which stands for live event assistant -- delivers personalized concert recommendations to users.
The original idea behind Lea was to build a conversational agent that users could chat with to explain what they're interested in, Canty said. But in the early stages of development, the team found that it wasn't able to build an agent that consistently delivered meaningful conversations using natural language processing (NLP) applications. Even something that worked most of the time wasn't going to be good enough.
"What we just kept coming into was the user expectations were exceeding the technology's abilities," Canty said. "If it's not right 99% of the time, the user gets frustrated."
Sometimes, a basic bot beats a virtual assistant
The team decided to dial back the scope of the project and focus less on NLP applications. Instead of learning user preferences via chat, users grant the bot access to their listening history on Spotify. The bot also looks at data from other sources, like Last.fm and Facebook's social graph.
An in-house data scientist built a machine learning model that uses this data to cluster users based on their interests. Through Facebook messenger, the bot delivers recommendations for concerts and other live events based on the machine learning model.
For now, the messages are one-sided; there's no chat functionality built-in. But, according to Canty, the information Lea is learning in its current state will form the foundation of a bot that is better able to have conversations with users, which is the future goal for Lea. He said more and more web traffic is going through smart speakers and other voice interfaces, so services of all kinds will have to get much better at having conversations with people in the future.
"The better we get [at] figuring out how to parse intents in [users'] natural language, the better we'll get at building skills for all those use cases," Canty said. "Even if it's not sexy NLP technology right now, there's still a ton of value in what we're doing."
NLP applications help drive insights
Alternatively, NLP applications can be implemented separately from bot technology. This is the approach New York-based Crisis Text Line is taking. The nonprofit provides intervention for people dealing with social and emotional problems for free via text messaging. The nature of the service means the organization has access to a large trove of natural language text data, said Scotty Huhn, data scientist at Crisis Text Line.
Huhn uses Python code via the Periscope Data analytics platform to analyze text data and look for correlations between specific words used by counselors and effective outcomes, as rated by the texters. He also groups counseling sessions by texters' issues to see what works best in different circumstances. The information about what works best is then passed on to the organization's counselors, and other data is published on the group's public website.
"We're really interested in the words and phrases that are highly correlated to effective counseling techniques," Huhn said. "We can make sure that counselors are always up to date on our learnings."
He's found that specific words often correlate with good outcomes. For example, he said he's seen a correlation between counselors telling texters that they were strong to reach out and effective outcomes. He said this kind of learning, based on NLP applications' technology, has made a significant difference in the ability of counselors to provide effective support to texters.
"We're able to extract these really powerful words," Huhn said. "These aren't just your standard conversations; they can get at the heart of experiences that people are digesting."