Facebook Launches Open-Source 'Blender' Chatbot Project for Developers and Innovators

Facebook's Blender: A Leap Toward Human-Like Chatbots

Virtual assistants on our smartphones are useful, but they currently lack the ability to engage in meaningful conversations. Designed mainly for simple tasks like checking restaurant reservations or providing weather updates, these assistants struggle with complex interactions. However, Facebook is making strides in chatbot technology with its recently launched open-source project, Blender.

Blender is touted as the most advanced open-source chatbot to date, boasting a remarkable 9.4 billion parameters—nearly four times more than Google’s Meena. Facebook's investment in Natural Language Processing (NLP) has clearly paid off, as the team behind Blender worked to overcome significant challenges in scaling its neural network. “As neural network models grow larger, they tend to perform better,” explained Stephen Roller, a research engineer at Facebook’s AI lab (FAIR). Training such a large model required innovative methods to distribute processing across multiple GPUs without sacrificing efficiency.

Yet, the size of the model is only one factor in its performance. Surprisingly, Roller noted that a smaller 2.7 billion parameter model slightly outperformed the larger version, emphasizing the importance of specialized skills in enhancing chatbot interactions. Blender was specifically engineered to exhibit empathy, personality, and knowledge, allowing it to respond appropriately based on the context of the conversation. This functionality is driven by Blender's unique Blended Skill Talk feature, which enables it to switch between different conversational styles fluidly.

Emily Dinan, another research engineer at FAIR, shared that their team has designed tasks to cultivate these conversational skills. “Our evaluation showed that models refined with targeted conversational datasets are not only more engaging but also feel more lifelike,” she stated. As a result, Blender can congratulate users on achievements or offer condolences—demonstrating an emotional intelligence that previous chatbots lacked.

In contrast to conventional assistants that provide surface-level information, Blender can engage in deeper discussions. For instance, if asked about Led Zeppelin, it offers comprehensive insights rather than just snippets from Wikipedia. This enhancement is achieved through pre-training using the Wizard of Wikipedia system, followed by fine-tuning with conversations between human pairs, helping to develop natural speech patterns.

Creating the high-quality conversational data needed for effective training is costly and challenging, which is why FAIR has developed a method to give chatbots distinct personalities. By utilizing character profiles—like that of a basketball enthusiast from Michigan with a family—Blender maintains consistency in its responses, reducing logical inconsistencies.

Feedback from users indicates that Blender surpasses other models. In a comparative evaluation using the ACUTE-Eval method, 67% of participants found Blender more human-like, and 75% preferred engaging in longer conversations with it over Meena. With these advancements, Facebook is setting the stage for truly conversational AI, paving the way for a future where chatbots can engage in more meaningful interactions.