Chatbots aren’t very good at small talk, but Facebook is trying to change that.
Artificial intelligence (AI) researchers at the social media giant believe that giving chatbots a ‘consistent personality’ might be the key to better conversation between humans and machines.
The team from Facebook AI Research recently released a paper which outlines their efforts to make the bots better conversationalists.
They say that common problems with chatbots when it comes to chit-chat include inconsistency, forgetting what people have said to them previously, and the fact that they’re “not very captivating“. They also they tend to fall back on responses like “I don’t know” when faced with a question they don’t understand.
To help solve these problems, the researchers have been giving chatbots a configurable but persistent persona, and training them to both ask and answer questions about personal topics.
Facebook has created its own dataset called Persona-Chat to train chatbots with. This contains more than 160,000 lines of dialogue between people who stuck to certain personalities when chatting to each other.
So far, the results are encouraging.
Interactions between chatbots trained on Persona-Chat data and real people didn’t score as highly on criteria like ‘fluency’ and ‘consistency’ as human conversation partners, but they were rated far better than those with bots trained on film subtitles.
The researchers hope that their findings will help train chatbots that can ask questions about users’ profiles, remember the answers, and use them naturally in conversation.
A challenge for future research is to find the right balance between fluency, engagement, consistency, and a persistent persona, the researchers said.