Amazon is working on a new training method to help improve Alexa's ability to answer more complex questions and solve your queries more usefully.
The company explained in a blog post how this new method works. In essence, the AI team is combining both traditional text-based search queries and custom-knowledge graphs to solve complicated questions.
Interestingly, these teaching experiments used several different search engines to start with an ordinary web search and extrapolate from there.
Standard algorithms are then used to pick out named entities and key elements within the results. From there a knowledge graph is created to help work out relationships between the data and assign "confidence scores" as well as identifying "cornerstones".
All this data is then whittled down according to the confidence score and whether the results relate to the cornerstone of the query. Everything that isn't relevant is removed, leaving only the most accurate answer to the initial query.
The company says that during the test its algorithm returned the correct answer and outperformed other state-of-the-art systems. The combination of both techniques apparently proving to be a lot more accurate than the traditional searches or systems that rely on either just a text search or just a knowledge graph.
The culmination of all this is the news that in future Alexa should become a lot more intelligent. You'll get a lot less "I'm sorry, I don't understand" responses and hopefully far more useful, accurate and correct answers to your queries.
