Google Has Made a New Chatbot Called Meena – Here’s What We Know

The new chatbot from Google has been trained using over 40 billion words and put them in social context for users.

Chatbots have been an intriguing part of our evolution, especially as our reliance on digital services continues to to increase.

However, this support system has largely failed to connect with users, failing to completely relate to conversations taking place. This shortcoming of chatbots (where conversation is over text messages) has kind of worked in favour of voice assistants to some extent.

Google has decided to take another stab at this space and has come out with its own chatbot called Meena, which is yet to make its public appearance. Early signs suggest the search giant is onto something that could change the dynamics of chatbots in the industry.

So what is so interesting about Meena, the chatbot? Here’s everything you need to know.

Google says Meena is an open-domain platform, which lets users type and ask queries from any platform and have a conversation without restricting its expertise. In a blog post Google says: “Meena has been put through an end-to-end, neural conversational model that learns to respond sensibly to a given conversational context.”

Using this model, Google researchers have tried to make sure Meena has a better grasp of predicting the next word that will be said by the user in a conversation.

Meena has a single Evolved Transformer encoder block and 13 Evolved Transformer decoder blocks. This might sound highly technical, but the below image shows the impact of using these many encoder and decoder blocks.

“The encoder is responsible for processing the conversation context to help Meena understand what has already been said in the conversation. The decoder then uses that information to formulate an actual response.” 

Google blog post

But talking about the early stage of its development, Google says after evaluating the risks and benefits of Meena, it will roll out in the coming months without giving a timeline.

“We are evaluating the risks and benefits associated with externalizing the model checkpoint, however, and may choose to make it available in the coming months to help advance research in this area.”

Google blog post

Behind the Scenes Effort

Compared to existing chatbots, Google says Meena has a good grasp on conversations, and tests suggest, its chatbot comes closest to having a normal chat. The below chart confirms Meena’s conversational skills are closest to how people converse on a daily basis.

But the work that has gone behind the scenes to get this conversion rate is mind-boggling. Google points out the Meena model has been trained with 40 billion words and 341 GB worth of text, filtered from public domain social media conversations. The model was trained for over 30 days (one month) on a TPU v3 pod which is basically a collection of 2,048 tensor processing units (TPU) cores.

So when can we expect Meena to become an intrinsic part of our digital lives? Even though Meena delivers on some aspects, Google says there’s a lot of work yet to be done before making it available to the end user through different services.