Competing with the global AI market, Google introduces Meena, a conversational agent that can make real conversation with humans. Meena is a conversational agent or a chatbot as people like to call it. It is a 2.6 billion parameter end-to-end trained neural conversational model that understands humans and can have a more sensible conversation with them. Google has the AI chatbot trained on 341 GB of public social-media chatter. And while it performs the tasks that regular conversational agents can do, it can also have a human-like open-ended conversation.
As more and more people opt for smartphones these days, it’s more likely that chatbots have become popular. It’s easy to command “Hey Siri” or “Alexa” and ask them to call someone we know or search the internet for us. And with better remodeling, the chatbots have, to some extent, made our daily lives more comfortable. However, the biggest concern Google has over these chatbots are how a lot of time they don’t make sense at all.
The inconsistency of the other chatbots in answering every question is what inspired Google to come out with Meena. According to its makers, Meena “can conduct conversations that are more sensible and specific than existing state-of-the-art chatbots.”
What Can Meena Do?
Google developed a new metric that captures the attributes for a more natural conversation. It is Sensibleness, and Specificity Average (SSA), and Google’s Meena scored a staggering 79%. The scores are pretty impressive compared to 86% scored by human conversation partners and 56% by Mitsuku, a state-of-the-art, four-times Loebner Prize winner chatbot. So, while most conversational agents would answer “I don’t know” or “that’s nice,” or something else. It might make sense; however, it isn’t specific. So, what Google aims to do with Meena is to be specific to the conversation. Hence, if you say to Meena, “I love EPL football,” the bot’s response won’t be “that’s nice.” Instead, it would say, “me too…I’m very fond of Manchester United.”
Meena has the Evolved Transformer seq2seq architecture, which helps minimize the perplexity or the uncertainty of predicting the next word in a conversation. It has a single Evolved Transformer encoder block and 13 Evolved Transformer decoder blocks. The encoder is liable for processing the conversation context, and the decoder is used to formulate an actual response. Both the tools help Meena understand the conversation and use the information to give out the best answer possible.
Google has yet to release the conversational agent to the general public. Users will only get a taste of the chatbot after it has vetted the model for safety and bias. However, learning about how much effort the company is putting towards the bot is undoubtedly impressive. Now that people have started using conversational agents to have an actual conversation to feel less lonely and anxious, maybe having a chatbot that is more sensible and understanding isn’t a bad thing. What’s more important is for Google not to repeat the mistakes made by Microsoft’s chatbot Tay.