Facebook AI and machine learning division FAIR blog announced that after years of research, it has built and opened up a new chatbot, Blender. It is also the largest Open-Domain chatbot ever built. “Today, we’ll release a complete model, code, and evaluation settings so that other AI researchers can reproduce this work and continue to advance the conversational AI study.” “
Blender was the first chatbot to integrate a variety of conversational skills, including empathy, knowledge, and personality, into a single system. According to human evaluators, it is better and more human in terms of engagement than others.
At the same time, Blender includes improved decoding technology, innovative skills fusion, and a model with 9.4 billion parameters, 3.6 times higher than the largest existing system, Google chatbot Meena. It also includes technologies that are equally important for mixing skills and generating in detail.
Facebook says the first step in creating a chatbot is mass training, with Blender using 1.5 billion conversations as training material. At the same time, the researchers introduced a new hybrid skill task (Blended Skill Talk, BST) to train and evaluate the model’s conversational skills, and Blender combined many of Facebook’s previous research findings.
BST includes the following skills:
Participate in using personality (people chat)
Participate in the use of knowledge (Wikipedia’s wizard)
Show empathy (sympathetic dialogue)
Seamless mixing of all three (BST)
The results of the test, which was unilaterally provided by Facebook, showed that 67 percent of evaluators said Blender sounded more human, and 75 percent said they would rather have a long conversation with Blender than with Meena.
However, Facebook also says that Blender chatbots still have many weaknesses over humans, and finding an assessment that better exposes those weaknesses is an open question and part of its future research agenda.
“Currently, we are exploring ways to further improve the quality of model sessions as you engage in longer-term conversations with new architectures and different loss functions. We are also focused on building more powerful classifiers to filter out harmful language in conversations. We have seen initial success in research to help alleviate gender bias in chatbots. “With open source fine-tuning and automated and manual assessments, we hope that the AI research community will build on this to drive dialogue AI forward.” “