Facebook’s AI Research team (FAIR) is teaching chatbots how to talk more like humans and engage in ‘small talk’ to ensure a lively conversation than what current chatbot models offer. According to The Verge, which has reported on this, FAIR’s paper on the issue is yet to be published, and it talks of solving the problem of chat models which lack personality. A PDF version of the paper is already available online.
According to the paper, there are several existing problems with the current “chit-chat models” of bots. This includes lack of specificity, lack of a consistent personality and the fact that these are often not “very captivating,” when it comes to conversations with users. The paper notes that currently, even a short conversation with the existing chatbots will “quickly expose their weaknesses”. Another problem with chit-chat models is the lack of “an explicit long-term memory” and a tendency to give “non-specific answers like “I don’t know.” The paper examines how they can make chatbots more engaging for users.
As part of the research, Facebook’s AI team has collected and trained data models based on profile conditions, which includes information about the person with whom the chat bot is talking. This helps to create a more engaging conversation. The Verge reports that Facebook’s teams relied on data-sets from Amazon’s Mechanical Turk marketplace, which included more than 160,000 lines of dialogue, in order to build and train these new personality-based conversation models.
Facebook’s paper also show these profile based chatbots will build up a persona, which is then used to carry the conversation forward. “This profile can be stored in a memory-augmented neural network and then used to produce more personal, specific, consistent and engaging responses than a persona-free model,” says the paper. The team saw that their new models, based on an assigned character or personality type resulted in more engaging conversation with users.
The paper also says, “We believe PERSONA-CHAT will be a useful resource for training components of future dialogue systems. Because we have paired human generated profiles and conversations, the data aids the construction of agents that have consistent personalities and viewpoints.” Facebook’s research comes at a time when interactions between humans and chatbots, virtual assistants continues to grow.
There are home speaker devices like Amazon’s Echo powered by Alexa, Google’s Home and Apple’s upcoming HomePod speaker; all three devices are powered by virtual assistants. Virtual assistants like Google Assistant, Samsung’s Bixby, Apple’s Siri are also present on major smartphones. For AI researchers trying to improve interactions between these assistants and humans will be a priority and the past few years, we have seen improvements in the field. For instance, Siri and Google Assistant can now understand natural language conversations, meaning a user does not have to rely on a fixed query phrase to get a particular answer. Facebook’s research could just take these conversations further, extending it to small talk as well.