Amazon’s Alexa voice assistant will get better at predicting your next request, thanks to new machine learning systems. The capability is already available to Alexa customers in English in the United States, and it brings Amazon’s voice assistant closer to more natural conversations. The company notes in a blog post that early metrics have shown that the new system is increasing customer engagement.
In a blog post, Amazon announced that Alexa’s new capability lets it infer the customer’s latent goals, which might not be expressed directly, but are implied within the request. For example, if a user asks how long does it take to steep tea, the next question might be about setting a timer for this task. With the new capability, Alexa could preempt this and answer that question, ‘Five minutes is a good place to start’, then follow up by asking, ‘Would you like me to set a timer for five minutes?'”
Another example given by Amazon is where a customer asks Alexa about the weather at the beach. Alexa might then guess that the user needs other information for planning a trip to the beach. Amazon says the goal for Alexa is that customers should find interacting with her as natural as interacting with another human being. In September, the company had announced “natural turn-taking” which allowed customers to talk to Alexa without having to say the wake word all the time.
The blog post explains that in order for such transitions to take place it relies on a “number of sophisticated algorithms,” which work to detect these “latent goals” and form them into actions. The idea is also to ensure that when Alexa offers these suggestions, they don’t feel disruptive, notes the blog post.
How will Alexa get better at predicting user intent?
Amazon notes that not all requests are are suited for this kind of tasks. For example, one earlier prototype would incorrectly ask users if they wanted to play chicken sounds in the follow up, when they asked about chicken recipes as the first request.
Amazon is using a “deep-learning-based trigger model” to help Alexa determine user intent. This model keeps in mind several other aspects of the sentence, such as the text and whether the customer has engaged with Alexa’s multi-skill suggestions in the past. If the model finds the “context” suitable, then Alexa suggests a skill to fulfil the latent goal or the one that is not expressed directly.
Amazon says these “suggestions are based on relationships learned by the latent-goal discovery model.” An example is that the model might have learnt over time that customers who ask how long tea should steep, and frequently follow up by asking Alexa to set a timer for that amount of time. Now with the new capabilities, Alexa will be able to make that connection and offer the timer suggestion on its own.
The latent-goal discovery model analyses multiple features of customer utterances, including mutual information. There’s also “semantic-role labeling model” at play, which “looks for named entities and other arguments from the current conversation, including Alexa’s own responses,” notes the blog.
Finally, Amazon also relies on “bandit learning,” where “machine learning models track whether recommendations are helping or not.” The post notes that “underperforming” ones are then suppressed.
📣 The Indian Express is now on Telegram. Click here to join our channel (@indianexpress) and stay updated with the latest headlines