Hacker News Reader

How do such personal agents work? I mean after the speech recognition and NLP (part-of-speech tagging, named-entity recognition, etc):

Do they use template databases for different topics? (like afaik WolframAlpha) Do they use (scientific) ontologies or are the templates more flat and stored in a SQL database? [and web search results as fallback]

View on HN

> Do they use template databases for different topics? (like afaik WolframAlpha)

What do you mean by a template database?

NLP sentence templates:

"How tall [object]?" / "[object] height"

"Becoming a [job]"

"How much is a [unit] of [object]?"

"How old is [object]"

"When is the next [event]"

My question is: Do Siri/Cortana/GoogleNow use such flat templates (per topic) or do the use an IR ontology? http://en.wikipedia.org/wiki/Ontology_(information_science)

We don't use flat templates in Alpha. We use a full context-free grammar that turns the question into a symbolic representation of what the user wants, which can naturally involve chaining. E.g. "when was the president of the US was born".

Then we execute that symbolic representation using a variety of strategies. There are funnily enough quite a few cases where we can understand the question but don't have the curated algorithms to actually answer it.

Reply on news.ycombinator.com