Meena: Google’s next virtual assistant has trained 40 billion words

Meena: Google’s next virtual assistant has trained 40 billion words

New York, 5.2.2020

Most virtual assistant vendors like Google and Amazon have spent the last few years developing more dialogue-oriented A.I. bots that go beyond the usual jokes and commands and can essentially talk to you like a real person. A Google announcement released today describes a „human-like“ chatbot called Meena that „can enter into a conversation on any topic“. What is it?

Meena is an open domain chatbot, unlike his colleagues, such as Google’s assistant. This means that Meena is not built around a limited set of data that is hand-picked for specific tasks. Instead, Meena is designed to be contextually relevant and to talk to you constantly – no matter what the topic is.

This is possible thanks to Google’s extensive data collection. The company claims that Meena has been „trained on 40 billion words dug up and filtered from publicly available social media conversations.

More importantly, by using Meena, Google avoids the most common mistake these wizards make – namely perplexity. The machine learning models that Siri and Google Assistant use cannot handle multiturn dialogues like humans. When in doubt, they simply say they can’t understand the query and retrieve a web result.

To ensure that Meena doesn’t fall victim to the same stumbling blocks, Google adds a second parameter to its algorithms that is not only smart enough not to answer with gibberish, but can also provide a specific answer.

„For example, if A says ‚I love tennis‘ and B replies ‚That’s nice‘, the statement should be marked as ’not specific‘. This answer could be used in dozens of different contexts. But if B replies ‚Me too, I can’t get enough of Roger Federer‘, then it is marked ’specific‘ because it relates closely to what is being discussed,“ Google added in the paper.

Based on these metrics, the average person reaches 86%. Google says it has so far managed to get Meena to an astonishing 79%. In comparison, Pandora Bots‘ A.I. agent, who has won the Loebner Prize for four years, an annual competition that rewards the programs that are most human, scored 56% in Google’s tests.

Google isn’t the only company that is working on human-like conversation bots. Through its acquisition of Semantics Machines, Microsoft has been working on the development of multiturn dialogues in chatbots for over two years. Samsung has announced that they will introduce a virtual human at this year’s Consumer Electronics Show.