IBM’s concierge bot serves hotel guests in interesting ways

Posted by:

I learned three things from IBM’s Connie the concierge robot when I visited her at the Hilton in McClean, Virginia. First, my Southern accent is much stronger than I realized. Second, voice recognition systems still have trouble understanding some of us. Third, the natural language processing capabilities of this Watson-driven robot are impressive, but our brief acquaintance indicates that they fall short of being cognitive.

According to IBM’s March press release about the project, “Connie uses a combination of Watson APIs, including Dialog, Speech to Text, Text to Speech and Natural Language Classifier, to enable it to greet guests upon arrival and to answer questions about hotel amenities, services and hours of operation.”

I asked Connie a series of questions about those subjects, and she seemed to be best at giving directional advice. For example, when I asked how to get into DC – about 20 miles away – from McClean, she gave me driving directions as well as two public transit options. Connie uses WayBlazer technology for travel advice, and that shines through here.

However, when I asked “What is a good place to eat?”, Connie responded with “What place do you mean?” My colloquial question clearly confused the robot, so I clarified with, “What is a good restaurant nearby?” In response, Connie suggested that I go to the hotel concierge since she didn’t know the answer to my question “yet.”

If you take a look at the accompanying video, you’ll hear that I asked Connie four times where the hotel pool was. When I first asked, she seems to have picked up on the “where” in my question, asking me again, “Which place do you mean?” On my second attempt, she said that she couldn’t hear me. On the third go ’round, she said she didn’t know the answer to my question. But on the fourth try, she answered it like a champ. Granted, my long vowels might have tripped up the bot, but “pool” seems an easy enough term to pick up on. If Connie needed the additional “hotel” modifier, I would imagine a cognitive system to follow my question with a “Did you mean the hotel pool?” Perhaps the ambient noise in the lobby or my accent interfered, but speech to text continues to be an issue.

IBM says that Connie will learn and adapt to guests’ questions, but I saw no way to help the robot learn. There was no obvious feedback method so that it could learn from guests’ responses to its answers, nor was I successful in asking it a follow up question to any of my first questions. As we have defined them here, cognitive systems’ answers should change with new information and be iterative and stateful. That is to say, they should remember previous interactions and return new answers that build on that new information.

IBM is surely collecting the questions guests ask, and I can see how Connie could be fed new questions and answers from that data collection. Along with Watson’s excellent natural language processing, this additional information will make guests’ interactions with Connie that much richer in time. But in the 10 minutes I spent with the robot, it did not appear to be learning – yet.

The hotel also is keeping a list of the questions guests ask Connie to mine the data for new ways to better serve guests. My questions will surely throw off that effort, but not as much as others. A hotel employee told me that people often ask the robot personal questions, seeming to think of it as a fortune telling robot instead of a service bot. He declined to say whether that annoyed any of the hotel staff – human or robot.


About the Author:

  Related Posts