Facebook has been working on chatbots for some time now. In particular, the conversational computer programs were a big focus during Facebook’s annual F8 conference last year.
Facebook has also featured a digital assistant called M in Messenger, although it ended support for M earlier in the month.
Now, Facebook researchers are working on improving how chatbots speak.
In a new report, the team identified some of the notable limitations with chatbot speech, including the fact that the AIs don’t have a consistent personality and don’t remember things they’ve said before, making dialogue with them feel unnatural.
The researchers also noted that some chatbots respond to many queries using generic pre-programmed responses or lines taken from famous movies, further weakening their ability to communicate effectively.
To improve all of this, Facebook has created a new dataset called Persona-Chat, which uses more than 160,000 lines of dialogue sourced from Amazon’s Mechanical Turk marketplace — the industry go-to resource for human data used to train AI.
Persona-Chat is evaluated using four criteria — fluency, engagingness, consistency and profile detection. With this in mind, Facebook works to improve chatbot speech to sound less robotic and more personalized. For example, instead of saying “I have been married four times and widowed three,” the Persona chatbot may say “I have a lot of experience with marriage.” Likewise, “I love the beach” may turn out as “to me, there’s nothing like a day at the seashore” with Persona.
Facebook researchers have also created short biographies for chatbots. One such background, according to The Verge, goes as follows: “I am an artist. I have four children. I recently got a cat. I enjoy walking for exercise. I love watching Game of Thrones.”
Chatbots surely have a long way to go before becoming completely human-like, but the Persona-Chat dataset is certainly an improvement.
Source: Facebook AI Research
The post Facebook researchers are working on humanizing chatbots appeared first on MobileSyrup.