And while that process appears straightforward to the user, Amazon insists that the actual backend algorithms that drive the feature are quite complicated. The trigger that activates the system and offers a latent goal — setting the reminder alarm or turning on the DVR — is trained on a deep-learning model that takes into account several aspects of the conversation’s context. What’s more, the model will improve over time as it associates more contextual clues with latent goals based on its continued conversations with users.
The system improves itself by actively learning, “which identifies sample interactions that would be particularly informative during future fine-tuning,” according to a Tuesday press release. The system also looks for “named entities and other arguments from the current conversation,” automatically labelling and formatting them for use by other Alexa skills. The most accurate and useful recommendations are implemented while the least are discarded, a process known as bandit learning.
It’s currently available in English for the US so developers can begin implementing it immediately.
