- Facebook launched a new chatbot ‘Blender’ as an open-source project and claims that it can converse about approximately anything interestingly and engagingly.
- Blender was trained on large training data models, and Facebook performed a test showing Blender is more engaging than Google’s chatbot ‘Meena.’
- Despite the impressive results, Blender still has some limitations, and Facebook is planning to create chatbots that can respond to visual cues too.
Chatbots and virtual assistants have made considerable progress in recent years. Yet, they are terrible conversationalists, and most of them tend to be highly task-oriented, while others are highly frustrating or boring. As these bots are becoming increasingly popular, the deficiencies solely grow more apparent.
Facebook has now launched a chatbot, named ‘Blender’ as an open-source project and claims it to beat Meena, Google’s chatbot as best in the world. Facebook states that Blender can talk about approximately anything engagingly. Stephen Roller, a Research Engineer at Facebook and co-leader of this project, said that if a person can solve dialogue, he has solved all of AI and vice versa. Facebook trained Blender on 1.5 billion Reddit conversations, which were publicly available to create an ability in it to generate responses in a chat. The Blender was then fine-tuned with three data sets to teach it empathy, knowledge, and personality. Blender is 3.6 times larger than Meena and requires two computing chips to operate.
Facebook performed a test in which 75% of human evaluators found Blender more engaging than Meena, 67% said it sound more like humans, while the Blender fooled 49% into thinking the conversation was more human than the chat between real people.
Still, Blender’s communication skills cannot be compared to human’s capabilities as the chatbot was evaluated only on short conversations. The team assessed the chats with 14 turns, and according to researchers, Blender will soon stop making sense if the team will keep chatting longer. Emily Dinan, the other project leader, stated that Blender is not able to recall conversation history after a few turns.
One of the limitations of this chatbot is that Blender can also make up facts, and the Facebook team plans to integrate a knowledge database with Blender’s response generation. Facebook may also face another challenge with this open-sourced chatbot, it may start saying biased or toxic statements during the conversation as these systems are trained on social media. To fix this issue, Facebook’s crowd workers filtered out toxic language from the three data sets (for fine-tuning) but could not do the same for Reddit data.
Facebook Artificial Intelligence team plans to create a sophisticated chatbot that will be able to respond to visual cues as well, such as it is already developing the ‘Image Chat’ system.
Read next: 'Distractions' keeps us - humans different from bots, reveals a Twitter study