Google's Conversational AI Can Converse Like A Real Human

Conversational AI is artificial intelligence that can speak like a real human or interpret human words. However, making such an AI is still far-fetched as no tech company has come close to making a model that can converse like a real human.

It is true that we have a virtual assistant that can answer our questions, control our home appliances like Siri, Google Assistant, and Alexa. If you ask, they will tell you jokes and respond to your questions even without a wake word. However, a conversation is still something only a human can do.

Google is the latest company to have progress in the endeavor to make a true conversational AI. Recently, it introduced Meena, an artificial intelligence system that has 2.6 billion parameters. One of Meena’s abilities is multi-turn dialogue and according to Google, the neural network is better than any other AIs in holding conversations.

With Meena, Google also introduces SSA (stand for Sensibleness and Specificity Average), a metric made by the company’s experts to measure the effectiveness of neural networks in maintaining conversations that make sense.

Real humans score around 86% on SSA while Meena is not so far behind with a rate of 79%. The holder of Loebner Prize in four years in a row, Mitsuku, a creation of Pandora Bots, scores 56%. Another example is XiaoIce, a Mandarin-Chinese-speaking AI of Microsoft got 31%.

Google used 40 billion words, a Transformer architecture’s variation, and a seq2seq model for its conversational AI. Transformer was released in 2017 but it took no time for it to become one of the highest-ranking language models.

SSA judges the ability to hold a conversation with a fixed prompt set. Each conversation needs to last more than 14 turns and does not exceed 28 turns. The results are the percentage of turns that the subjects are able to give a sensible response. If the answer is considered generic, the system will be penalized.

Google said in a blog post that in the next few months, it would make Meena available to researchers, but it would also try to avoid releasing a demo immediately.

Google’s standards to assess a conversational AI is different from those of other organizations.

The race to create an artificial intelligence system that can hold a real conversation is attracting the attention of many. Alexa prize is an annual challenge for students to encourage them to create AIs that can converse for up to 20 minutes.

Last year’s winner could hold a conversation for 10 minutes. This year, the latest round will take place in May. If you want to speak to the latest winner, all you have to do is to say “Alexa, let’s chat.”

Amazon is also interested in multi-turn dialogue as it is developing its own offerings as we speak.

Another tech giant that is investing in this field is Microsoft. With its acquisition of Semantic Machines in 2018, Microsoft has had the first success in making conversational AI last year to use in its Bot Framework.

As stated by Ashwin Ram, director of Google Research and former head of Alexa Prize, in 2017, a neural network that can converse like a real human can establish a stronger bond with the users while helping to eliminate loneliness.

>>> AI Discovered The Coronavirus Breakout In China Way Before Official News From WHO