Meena: Google's AI Agent Built for Human-Like Conversation
Thu, April 22, 2021

Meena: Google's AI Agent Built for Human-Like Conversation

The odds that we have engaged with a chatbot these past few days or weeks is high, considering the rising adoption of this technology. While most of us think that chatbots are only recently developed, the first instance of this technology goes back to the 60s / Photo by anyaberkut via 123RF


The odds that we have engaged with a chatbot these past few days or weeks is high, considering the rising adoption of this technology. While most of us think that chatbots are only recently developed, the first instance of this technology goes back to the 60s. During the mid-60s, Joseph Weizenbaum from MIT's Artificial Intelligence Laboratory created the first example of a chatbot, codenamed “ELIZA.” The chatbot was able to simulate computational understanding without having machine learning capabilities. 

Today, with artificial intelligence and machine learning, chatbots are much better and more sophisticated. They are easy to build and navigated by predefined flows and effectively perform. Thus, it’s not surprising that industries and companies are heavily investing in them. Forbes, a global media company focusing on business, investing, technology, entrepreneurship, leadership, and lifestyle, reported that chatbots are four times more effective at selling products than inexperienced human workers. 

A 2019 survey reported that twice as many consumers would knowingly engage with chatbots because they are “very helpful,” 83% of consumers said they’d make messaging their primary means of contacting customer support if they could be guaranteed an immediate response—up from only 76% in 2018, and 54% of consumers stated that customer service has gotten better in the past year—up from 43% in 2018. 

The State of Service released by Salesforce, an American cloud-based software company headquartered in San Francisco, California, showed that the most common uses of AI chatbots include providing self-service in simple scenarios, gathering initial information about a service case before handing off to an agent, providing agents with guidance and recommendations as they handle cases, and greeting customers when they call. Aside from that, Salesforce also revealed that 58% of customers stated that emerging technologies such as chatbots have changed their expectations of companies and 77% said that chatbots will transform their expectations of companies in the next five years. 

Just like other technologies, chatbots are getting more and more sophisticated. Conversational AI, the set of technologies behind automated messaging and speech-enabled applications that offer human-like interactions between computers and humans, is something that companies and customers are investing in. It can communicate like a human because it can recognize speech and text, understand intent, decipher different languages, and respond in a way that mimics human conversation. 

What is Conversational AI?

Conversational AI is the mature version of chatbots. It can perform multi-turn conversations and execute judgment-intensive tasks just like humans. Since chatbots have failed to deliver on expectations, the enterprise industry has turned to conversational AI to be used in complex areas such as banking, insurance, and telecommunications. 

MarketsandMarkets, an online site that provides quantified B2B research on 30,000 high growth emerging opportunities/threats that will impact 70% to 80% of companies' revenues worldwide, projected that the global conversational AI market size will grow by $15.7 billion by 2024 from only $4.2 billion in 2019 at a CAGR of 30.2% during the forecast period. The growth is driven by the increasing demand for AI-powered customer support services, omnichannel deployment, and reduced chatbot development cost.

There are several things that separate a chatbot from a conversational AI. According to a developer and provider of AI-based workplace communication platform, a conversational AI is capable of understanding complex sentences of human speech in the same way humans do, learning the history available from existing enterprise interactions, referencing previous transactions made by a person and trying to fix things, and supporting sophisticated enterprise security considerations. 

Google’s New Conversational AI

The most popular AI assistants that we have today such as Alexa, Google Assistant, or Siri aren’t that conversational. While they can answer questions and tell jokes, they can’t communicate as humans do. To address this, Google introduced an AI project called Meena that aims to develop the first digital assistant that can truly hold a conversation. In a paper titled “Toward a Human-like Open Domain Chatbot” published on preprint repository arXiv, the Google team stated that they wanted Meena to bring its conversational ability close to that of humans. 

Meena, a neural network with 2.6 billion parameters, was trained on 40 billion words and utilizes a seq2seq model and a variation of the popular Transformer architecture. Transformer, which was first released by Google in 2017, is widely acknowledged to be among the best language models available. Aside from the conversational AI, the tech giant also introduced a metric called Sensibleness and Specificity Average (SSA), which measures the ability of agents to maintain a conversation. It analyzes and evaluates the conversations depending on static performance with a fixed set of prompts or interactive performance.

The team reported that each evaluated conversation in SSA was required to last at least 14 turns but no more than 28 turns. After that, the results arrived at based on the percentage of turns considered specific or sensible. In initial tests, Meena scored 79%, which was higher compared to Mitsuku, a Loebner Prize-winning AI agent developed by Pandora Bots. It only scored 56%. According to The Verge, an American technology news and media network that publishes news items, long-form feature stories, guidebooks, product reviews, and podcasts, conversational AI like Meena that can maintain a conversation may be able to secure closer bonds with humans and do things like providing emotional support to people. 

At the same time, Meena addresses the problems with common voice assistants today. Most of them handle multi-turn dialogues like humans. “For example, if A says, ‘I love tennis’ and B responds, ‘That’s nice,’ then the utterance should be marked, ‘not specific.’ That reply could be used in dozens of different contexts. But if B responds, ‘Me too, I can’t get enough of Roger Federer!’ then it is marked as ‘specific’ since it relates closely to what is being discussed,” Google said in the paper. 

Unfortunately, Meena likely won’t make it to your devices anytime soon. However, the tech giant assured that voice assistants will undergo dramatic, fundamental upgrades in the next few years. Google is not the only tech company that’s working to develop more conversational AI platforms. For instance, Amazon is starting to grow its multiturn dialogue offerings and Microsoft is also beginning to showcase more multiturn dialogue for users of the Microsoft Bot Framework.

Google introduced an AI project called Meena that aims to develop the first digital assistant that can truly hold a conversation / Photo by Ken Wolter via 123RF