Meet Meena: Why you'll want to hire this Google chatbot

Google reports a breakthrough in chatbot technology -- it's the perfect idea for replacing its own Google Assistant ... and search. (And read on to judge Meena's joke-telling ability.)


At some point in the future -- it could be 25 years from now, or just five years, most of our interaction with computers and the internet will happen through spoken conversation.

This is true not because any particular technology is inevitable. It's true because user interfaces always evolve toward what's easiest for human users. People are hardwired for spoken language, and so ever-increasing compute power and software innovation will have to be applied to making computers talk.

But that's harder than it sounds. It turns out that in order for a machine to carry on a plausible conversation, it needs "knowledge" of the world. To keep up with human conversation, a machine needs the capacity to reason, to understand context, to be creative and to make sound judgements about a thousand different things, including relevance.

In other words, human speech isn't about spitting out words. It requires something approaching a human brain.

That's why it's astonishing that one of the biggest apparent technology breakthroughs of the last few years has been announced, and hardly anyone is paying attention to it.

The meaning of Meena

A chatbot is simply a software program you can hold a conversation with.

Most commercial chatbots deployed by large organizations are designed for narrow uses, such as customer service. These narrow-function chatbots are called closed-domain chatbots. Meena is an example of an open-domain chatbot -- one designed to converse on any topic that can function as a "friend," advisor and even teacher. An open-domain chatbot needs the knowledge and capabilities of thousands of closed domain chatbots combined.

Google unveiled this week an open-domain, neural-network-powered chatbot called Meena, and claimed that it's the best chatbot ever created. There's good reason to believe this claim is true. (Google declined an interview for this article.)

Meena is based on new technology, old technology, new approaches and mind-blowing quantities of data. Researchers fed Meena 341 gigabytes of social media conversation from public social media posts. It has 2.6 billion parameters -- far more than other leading chatbots. The dataset is filtered through, among other things, an algorithm that removes offensive content.

[ ICYMI: Will Chromebooks rule the enterprise? (5 reasons it may)] 

Google says Meena is designed to be specific, which would be impressive, and sensible -and mind boggling astonishing.

Google has invented a new metric to keep Meena from going off the conversational rails, as most chatbots have traditionally done. It's called the Sensibleness and Specificity Average (SSA) metric, and it judges whether each word makes sense within the context of the whole thread of conversation, rather than as an isolated response to the previous user input.

Conversational chatbots have been around for decades. They rely on tricks, such as generic vagueness in response to sentences they don't understand.

When someone says something to a chatbot that it doesn't understand, it's called perplexity. So part of the parlor trick with conversational agents is the graceful handling of perplexity. For example, if you tell a typical chatbot: "I like to scuba dive," the response might be: "I'm glad you like to scuba dive." It's a plausibly human-like response, but it's obvious that the chatbot is exercising this fallback option: Just say you're glad, followed by whatever they said. More importantly, the response is useless. That's why most chatbots are novelties and parlor tricks, rather than useful conversational agents.

Meena's specialty is the minimization of perplexity itself, rather than focusing on how to convincingly hide the perplexity with general or all-purpose responses.

[ More on chatbots: Is that you or a virtual you? Are chatbots too real?]

Meena scored 79 percent on the SSA. That's lower than the average human score of 86 percent, but much higher than the highest score of the previous Loebner Prize chatbot champion, Mitsuku, which scored 56 percengt. (You can chat with Mitsuku here.) In other words, Meena is theoretically closer to the ability to converse to humans than to the second best chatbot. Google researchers claim that human-level SSA is "within reach."

To be clear, these are claims, not facts. Until we can try Meena for ourselves, we're taking Google's word for it. The SSA is Google's own benchmark. And all the judgments about Meena come from its own creators. (Google may demonstrate Meena, or even make it publicly available, May 12, at Google's I/O developer's conference.)

Still, the claims are credible -- as well as incredible, in the sense of being extraordinary.

How good chatbots go bad

Google has not released a demo version for public use. The company plans to first make sure Meena is safe and unbiased. Good thinking, Google.

Four years ago Microsoft unveiled a chatbot called Tay, which was designed to absorb the language of the people who interacted with Tay on Twitter. Within 24 hours, trolls flooded Tay with the language of racism and misogyny, which turned Tay into a woman-hating racist. Garbage in, garbage out.

Microsoft was embolden to introduce Tay by the success of its Chinese-language Xiaoice chatbot, which was launched in 2014 and which has more than 660 million users. As with Tay, Xiaoice was equipped with the capacity to parrot social media chatter as a shortcut to natural language responses. The difference is that Chinese social nets are censored by the Chinese government, so the input came pre-sanitized.

Recently, Xiaoice has taken a dark turn. Microsoft launched a week-long trial Jan. 22 to test 999 individually customized "virtual girlfriends" based on Xiaoice with the same number of beta testers. The idea was to test Xiaoice's utility as an emotional companion.

Google appears to be trying to avoid the dark alleys of Tay and Xiaoice, and instead create a chatbot that's interesting, useful, enjoyable and even enlightening.

How Meena could affect your career and your life

One of the most culture-shifting technologies introduced in the past decade is the general-purpose virtual assistant. (Apple acquired Siri in 2010.) This was kicked into high gear with the mainstreaming of the smart speaker. (The Amazon Echo first shipped in 2014.)

The most popular Virtual assistants, including Apple's Siri, Amazon's Alexa and Google's Assistant, can't hold general conversations. They can reliably set a timer, turn off the smart lights and tell you the weather. But if you ask them something beyond their narrow programming, they'll tell you they don't know or they'll show you a website or search result. And some of them offer a small repertoire of canned, prefabricated quips and humorous responses.

Meena is designed to converse more "intelligently." Instead of tricking you with clever ways to answer questions it doesn't understand, it aims to actually understand the questions. It's designed to maintain the context of those conversations. And it's designed to be creative.

One of the most astonishing feats performed in the research is that Meena invented a joke.

Here's the whole conversation in which that invention occurred:

Human: I like how cute cows are
Meena: Ah, interesting. You like how they are cute, right?
Human: Yeah, and they’re really smart. 
Meena: I heard they go to college
Human: Cows go to college?
Meena: I heard that a cow went to Harvard.
Human: What did the cow study?
Meena: Bovine sciences.
Human: Do horses go to Harvard?
Meena: Horses go to Hayvard.
Human: That’s a pretty good joke, I feel like you led me into it
Meena: You were trying to steer it elsewhere, I can see it

There are several amazing things happening in this thread.

Specifically, the joke "Horses go to Hayvard" is the invention of Meena. A "dad joke," to be sure, but very human and pretty funny.  What's human about it is that the joke itself was not in the data. Meena coined the word "Hayvard" for the purpose of humor.

On a practical level, Meena "steered" the conversation toward this bit about cows going to Harvard when the subject of cows came up. This comment is based on Meena's knowledge of the world. In fact, a retiring Harvard professor 11 years ago brought a cow to Harvard as a humorous stunt.

[ Related: How AI is revolutionizing financial services]

Even more impressive, Meena introduced this knowledge to the conversation vaguely and casually, but could have presumably answered any questions about it. However, the context of the conversation was jokes and puns, and so Meena kept it vague and light, as a human would.

This and other examples in the published research paper suggest that a Meena-based assistant could engage with you in free-form, exploratory conversations that are intellectually and professionally enriching. It would be like having an assistant with a photographic memory who had read millions of articles, books, posts and other content.

Going even further, it would be trivial to plug company databases, intranet pages, performance data, sales figures and even all Slack chatter into the Meena knowledge base for the exclusive use of authorized executives like you. The result would be a conversational agent that could surface insights, give facts on demand and basically serve as the ultimate business tools.

Data access is the easy part. Technology that can talk based on data is the hard part. And it looks like Google is "steering" us in that empowering direction.

Copyright © 2020 IDG Communications, Inc.

It’s time to break the ChatGPT habit
Shop Tech Products at Amazon