AI Daily - Artificial Intelligence News

View Original

Facebook's AI Chatbot Wants to Talk With You

Facebook AI recently announced its state of the art open-source chatbot Blender. In their blog post, they say it is the largest open-domain chatbot, which means it can chat about anything and is "more human" than its predecessors according to human evaluators. It is the first chatbot to combine empathy, knowledge and personality amongst other conversational skills into one. When compared with Google's latest Meena chatbot, 67 percent of human evaluators said that Blender sounds more human and 75 percent said that they would rather have a (long) conversation with Blender.

An example conversation with Blender (Source: Facebook AI)

One reason Blender is better than other chatbots is due to large scale training. Facebook AI used large Transformer neural networks that had up to 9.4 billion parameters and pretrained them with a vast amount of conversational data. This is 3.6 times greater than the current largest system. Something to note is that even though the chatbot is trained using so much data, it does not necessarily mean that it will learn the traits of the best conversationalists. It can in fact "make the model imitate poor or even toxic behaviour". This is why Blended Skill Talk (BST) was introduced for training and evaluating purposes regarding these desirable conversational skills. This involves blending three different skills: personality, knowledge and empathy. Even though this is a difficult task, it is vital in the skill for a chatbot to notice when someone's tone changes from joking to serious. When the BST dataset was used to fine-tune the AI model, it had a profound impact on the chatbot's conversational ability.

To get to this stage, Facebook AI spent years improving each generation of their chatbots. They say that "true progress in the field depends on reproducibilty" and as a result they have made the Blender chatbot publicly available. The research paper is published online and they have also made the code available so that AI researchers can continue to improve on the work done by Facebook.

(Thumbnail Source: Unsplash/Austin Distel)