Google’s next big thing in the market is an AI-based bot named Meena that is expected to be far superior to the existing voice-based bot applications like Alexa and Siri. This information has been shared by a couple of research scholars with Google Research in a blog post.
The researchers claim that the main problem being faced by people using the bots is that the bots don’t comprehend many conversations and therefore lack the ability to respond appropriately. The underlying confidence Google has in Meena is that it is backed by a neural network of 2.6 billion parameters. This, they claim, provides it enhanced conversation skills.
The application of artificial intelligence with the larger neural network ensures that the bot figures out the context of a conversation and does not leave the user in any confused state on the response.
From the technology perspective, Google’s research team says it has used the Evolved Transformer seq2seq architecture. The essence of this architecture is that it improves the perplexity. To explain it in more simpler terms, the blog post again reiterates that the lack of conversational abilities has been an issue repeatedly faced with the bots.
To remedy this, Meena has been trained through 40 billion words which included social media conversations. The volume of data used for this training was 341GB.
There are blocks of encoder and decoder that assist the bot in deciphering the command given to it by the users and then in framing the response. Google says, in Meena, there is one evolved transformer encoder and 13 evolved transformer decoders for these functions.
The bot is not yet available on open source and there is no hint from the blog post when the bot Meena will be available on open source.