Amazon engineers want Alexa to be able to predict a user’s next question

Just a heads up, if you buy something through our links, we may get a small share of the sale. It’s one of the ways we keep the lights on here. Click here for more.

Amazon engineers continue to make Alexa a bit more interactive, smarter, and more natural. 

The latest machine learning update is in that general direction as it gives Alexa the capability to provide a suggestion based on your current conversation. Based on the given topic, Alexa will now try to predict what you might want next and suggest it before asking for it. 

For example, let’s say you ask Alexa, “Who’s the next guest on Saturday Night Live?” or “When’s the next FC Barcelona match?” In addition to telling you who will host the next SNL and when FC Barcelona will have their next game, Alexa might offer to set a reminder or set the DVR if you can’t catch the match live. 

On the surface, this seems like a straightforward process for the average Alexa users, but a lot goes into it

However, Anand Rathi and Anjishnu Kumar, engineers at Amazon, explain that this feature is driven by a complex, multilayer algorithm that considers a wide range of factors.

According to the recent press release, Alexa first decides whether it will try to anticipate your latent goal or not. To that end, Alexa relies upon a deep learning model that understands dialogue context. Furthermore, Alexa will use its past data to see if the user engaged with its suggestion tools in the past. 

In case the algorithm finds the context suitable, a suggestion will be made to the user. That’s what the engineers in Amazon call the trigger model. 

Then there is something that Amazon engineers call Latent-goal discovery. That is a deep learning model that considers various customer speech cues. For example, it sometimes analyzes the common nominator between the latent prediction and the immediate goal. As time passes, the latent goal discovery model will actively learn through the many interactions. 

The semantic-role labeling model is another integral part of this new feature. It looks into the named arguments and entities in the given interactions. It is a model that doesn’t just analyze the user input but also Alexa’s response.

Last but not least is bandit learning. Its purpose is to determine whether Alexa makes helpful recommendations or not. Those suggestions deemed as underperforming are immediately dismissed. 

So far, this new feature is only available in English and only to US-based Alexa users, but we can assume that as the systems in place become smarter, a larger rollout could be expected.

What do you think? Would you like Alexa to be able to predict what you might ask next? Let us know down below in the comments or carry the discussion over to our Twitter or Facebook.

Editors’ Recommendations:

• UK shoppers who bought the PS5 from Amazon are getting boned
• Amazon now has an online pharmacy with free 2-day shipping for Prime Members
• You can now ask Alexa to print things out for you
• Amazon just added a ton of new Alexa features – here’s how to use them

Deep Ballu