Amazon makes Alexa more talkative with new approach to deep learning

Get a free Techzine subscription!

Amazon has developed a new approach to deep learning, which has already produced new opportunities for speech and image recognition. The approach should help developers to more easily develop more natural speech experiences.

In addition, the approach, Alexa Conversations, should require a third less lines of code than the current approach, and require ten times less data to train the algorithmic models, writes Silicon Angle. The result of Alexa Conversations is more natural and flexible dialogues within individual skills, as Alexa’s capabilities and apps are called. In addition, it must be possible to use several skills fairly seamlessly in a single conversation.

Demo

The new features are already available as a developer preview, and were demonstrated by vice president and chief scientist for Alexa Rohit Prasad at the re:MARS conference. The new features include a cross-skill action predictor.

In the demo, a woman asked Alexa which films in the neighborhood were screening, and Alexa suggested booking tickets and easily dealing with the desired movie times. Alexa also asked the woman if she wanted to book a restaurant nearby, gave Alexa options for Chinese food, made a reservation on OpenTable and asked if the woman needed a taxi or Uber. The transport was then also booked.

According to Prasad, this is a big step for customer-oriented AI. Previously, it would have taken 48 requests from the user to get all this done, whereas now it only cost 13 requests. This means that Alexa will work again, instead of the user.

Operation

This is all possible via an artificial intelligence driven “dialog manager”, with an advanced dialogue simulation engine that automatically generates synthetic training data. By using APIs provided by developers for a skill, along with annotated pieces of dialogue containing requests for Alexa and actions likely to be undertaken by customers, Alexa Conversations generates dialogue flow and variations through a recurring neural network.

The network takes the history of the dialogue with it and predicts the optimal next action or step in that dialogue. The accuracy has to be increased and the work on design and code has to be reduced, says Amazon.

This news article was automatically translated from Dutch to give Techzine.eu a head start. All news articles after September 1, 2019 are written in native English and NOT translated. All our background stories are written in native English as well. For more information read our launch article.