The next time you’re hired, you might find yourself getting information about payroll, vacations, and expenses by talking to a chatbot instead of consulting a handbook for new employees or talking to someone in HR.
A startup called Talla, based in Boston, is working on chatbots designed to help new workers get up to speed and be more productive. The company is using advanced machine learning and natural language processing techniques in an effort to create software that is smarter than the average bot.
Talla recently launched a simple prototype bot for managing to-do lists on the workplace communications platform Slack. So far, about 600 companies have added the chatbot to their Slack channel and are using it, says May. A version of the chatbot will also be launched in coming weeks on HipChat, a competitor to Slack.
And, come October, Talla will be updated to include much more sophisticated HR capabilities. “If you came to work for us, rather than give you a whole stack of information, we’d say ‘Here’s a bot, Talla, and she’s going to walk you through it,’” says Talla CEO Rob May.
May says the vision is for Talla to grow from a to-do-list helper to an all-around workplace assistant. Talla plans to offer a free version of its technology, and to charge for more advanced versions. “Every day she’s going to tell you what you should be focused on; she’s going to ask you questions and give you feedback,” he says. “And if you have questions, ‘Hey, Talla, when do we get paid?,’ ‘Do we have Friday off?,’ ‘How do I change my dental?’—those are all things the bot will be able to do.”
Though the first chatbots were developed decades ago, an ongoing AI boom has renewed interest in them and kindled hope that a new generation of bots could be far more intelligent and useful. All sorts of companies are launching bots on platforms like Facebook and WhatsApp, and the popularity of workplace chat systems like Slack means that more chatbots could soon find their way into many office spaces.
Microsoft has committed to supporting the development of chatbots on its software, and Apple has announced that it will allow third-party developers use its voice-controlled personal assistant, Siri. Gartner, a company that provides market analysis, predicts that by 2019, 25 percent of homes will use smart assistants to access different online services.
The big challenge with all chatbots, however, is ensuring they understand what people mean when they say something. There has been dramatic progress in many areas of AI of late, but language understanding remains a huge challenge (see “A Tougher Turing Test Exposes Chatbots’ Stupidity”).
Talla is using some of the latest machine learning and language processing techniques in an effort to build a more intelligent system. For instance, its technology uses a deep-learning classifier—a large and somewhat crude network of mathematically simulated neurons that can be trained to recognize input—to determine whether a message is a question or a command.
Talla also uses “word embeddings”—a way of representing the meaning of words and phrases using mathematical vectors with many dimensions—to identify the meaning of a command or question, even if it is phrased in an unfamiliar way. May says the basic version of Talla is roughly 97 percent accurate at understanding commands that it should be able to carry out. Talla trained the to-do app using its own data. A company would most likely need to feed the system information from an HR handbook and might need to perform some training and testing.
Noah Smith, an associate professor at the University of Washington who develops computer algorithms designed to understand language, says that deep learning and word embeddings are increasingly being used to create natural language processing systems.
Smith says an HR chatbot “seems interesting and worth a try,” but adds that it could be difficult to pull off a comprehensive one because of the complexity of the language required. “Recruiting seems to me like a hard problem for a machine, since most models of language generation don’t yet integrate everything humans know and do to be persuasive, but it could be an exciting challenge,” he says.
Smith says natural language systems are generally improved through continual usage, and the current interest in chatbots could feed into that. “I think it’s exciting to see so many labs and companies exploring them and consumers actually using them,” he says. “I think they’ll provide a great platform to explore new capabilities and find out what people want, what’s easy to do, and what needs more research investment.”
It is certainly getting easier to build a chatbot to perform specific tasks. Dharmesh Shah, cofounder and CTO of Hubspot, a marketing software platform, is developing a chatbot for marketing professionals. His software, called GrowthBot, connects to a number of online services and uses natural language processing technology provided by Wit.ai, a company acquired by Facebook in 2015. But Shah says his bot doesn’t try to do anything too sophisticated.
“It provides a quick, convenient, and natural interface to what marketers need to do every day,” Shah says. For instance, a user can ask GrowthBot, “How was traffic to the website last week?” or “How was organic traffic to the site last month?” to retrieve information from the company’s analytics software.
May says that Talla might eventually help decide who should be interviewed for a job opening. Engineers at his company have designed a machine learning system that looks for similarities between the resumes of prospective hires and of existing employees who have proved successful. “It works really well,” he says. “It’s pretty cool.”