Although the process may seem simple to users, Amazon insists that the actual back-end algorithm that drives this feature is very complicated. The deep learning model is trained to trigger the system and provide triggers for potential targets (set reminders or turn on DVR). The model considers multiple aspects of the conversation context. More importantly, as the user continues to dialogue, the model will associate more contextual clues with potential targets, and the model will improve over time.
According to Tuesday’s press release, the system improves itself through active learning, “learning can identify sample interactions that are particularly useful in future fine-tuning processes.” The system will also look for “named entities and other parameters in the current conversation,”
There is currently an English version in the US, so developers can start implementing it immediately.