Hacker News new | ask | show | jobs
by captainmuon 3542 days ago
What I don't understand is why we don't try to solve this using a huge fact database derived from natural language parsing of the web. Basically the same approach as deep learning - something that was unfeasible 20 years ago, but due to massive improvements in processing power now works.

What I'm thinking of is basically SHRDLU [1] on steroids. Parse a ton of web pages. There are great natural language parsers that can parse most well-formed English sentences. Start with simple sentences like "Golden delicious is an apple" or "Barack Obama is the President". Then you store this in a Subject-Verb-Object database (I just learned that this is called a Triplestore [2]).

Every statement gets a plausibility value. Deal with ambiguity by adding multiple interpretations of a sentence (with different plausibilites if available). Assign an origin (e.g. website, author, quoted person ...) to each statement. Then, you could query this by asking "What do mice like?"... and it would make "Subject: Mice, Verb: like (enjoy), Object: ???" and return a list of solutions, ordered by plausibility.

Does anybody have any insight into why this isn't done or wouldn't work? It seems wierd that I can't ask my phone simple facts about the world, other than those whose form have been hardcoded.

(Now that I think about it, the opposite would also work. Hardcode a ton more commands. Hire 100 people, let them sift through the most common queries. Watch a few dozen testers, add add all queries they try to use. Instead of throwing computing resources at the problem, this would throw cheap labor at it.)

Anyway, it boggles the mind that I can't shout "OK Google, play 'itsy bity spider' on youtube" when my toddler demands it but won't release my phone :-). It opens a search and shows what I want as the first result (probably customized from my history), but I have to go the last mile myself.

[1] https://de.wikipedia.org/wiki/SHRDLU [2] https://en.wikipedia.org/wiki/Triplestore

4 comments

Obviously there's still a long way to go, but yes, Google is working on everything you suggest and more.

This is actually Google's bread and butter, and why Mossberg found so many cases where Google Now does better than Siri. It's not clear how Apple can catch up, either, given Google's massive advantage in training data.

It's not as easy as you think, but people are working on it. For example:

https://en.wikipedia.org/wiki/YAGO_%28database%29

Triplestores are defiantly cool; I used one once where I put in our user's birthdays and took a public ally available TS dump (I think it was from dbpedia) of celebrities and matched them up saying things like "hey you share a birthday with X"

What was cool about that was that I didn't have to know how to get their info nor what format it was in, provided that they followed some standard namespaces.

Look up JSON-LD while you are at it. This method can be extended to the IoT and then we will get some interesting outcomes.