Facebook AI open sources a chatbot with empathy and personality
The most recent addition to Facebook AI’s open source projects is Blender, a state-of-the-art chatbot. What sets it apart from other chatbots is its novel blending of skills, including empathy and personality. Let’s take a closer look.
Facebook AI has announced the open source release of its chatbot Blender in a blog post. This open-domain chatbot is the largest to date, consisting of 9.4 billion parameters, and it boasts a unique blending of skills.
According to Facebook AI, chatbots are generally designed for specialization in a certain area. What sets Blender apart is that it combines several conversational skills, for example, the abilities to ask and respond to a wide range of questions and to show empathy. Assuming a persona and being serious or fun have been added to the mix as well.
On Twitter, you can see an example conversation:
Today we’re announcing that Facebook AI has built and open-sourced Blender, the largest-ever open-domain chatbot. It outperforms others in terms of engagement and also feels more human, according to human evaluators. https://t.co/TdNxJpL0JI pic.twitter.com/1wo3ZxwDeU
— Facebook AI (@facebookai) April 29, 2020
Blender “feels more human, according to human evaluators,” as the blog post reads. Compared to dialogues created with Google’s latest Meena chatbot, which has not been publicly released though, 67% of human evaluators found that Blender sounded more human, and 75% would prefer to have a long conversation with Blender rather than with Meena.
Facebook AI has released the complete model, code, and evaluation set-up for Blender, which is available as 90M, 2.7B or 9.4B parameter version—the latter being 3.6x larger than the largest existing system. On the project’s website, you can also find several more example conversations.
In order to interact with the models, there are different hardware requirements: the 2.7B model needs a 16gb P100 GPU or better whereas the 9.4B model needs at least two 32gb V100 GPUs.
See the blog post for further details.