I agree I oversimplified. In the heat of the moment I was amazed by the fact that I checked one after another of the questions, and the response was right there in the title of the first result. It shouldn't take 4 years and millions of dollars to get from that to winning jeopardy (given also the ~7 seconds of advantage Watson get by getting the answer in the second it shows, while humans must listen to the host and read the question (I can't read it faster when he is reading it out loud.. Maybe Ken can..)).
> That's the tricky part, understanding what is being asked for.
That's right. Don't forget that Watson was programmed for Jeopardy, not generic natural language understanding. I am pretty sure that by analyzing past questions Watson has so many patterns hard coded in it, so it can quickly devise what kind of information it is looking for, and cross it with a few other search results.
> try typing a Jeopardy question into Google sometime.
Have you tried it? Most questions that Watson got right have the result in the title of one of the first three results (don't forget to add -jeopardy, to remove the noise from jeopardy archiving sites.
You're example is a hard one, but most others are not so hard.
Here are a few: "Elected every 5 years, it has 736 members from 7 parties" First result - "The European Parliament" - take that verbatim and you win.
"While Maltese borrows many words from Italian, it developed from a dialect of this Semitic language". First result - "Arabic Language".
"Gambler Charles Wells is believed to have inspired the song "The Man Who" did this "At Monte Carlo" Second result - "The Man Who Broke the Bank at Monte Carlo (song)". Cross that with the "this" in the question (familiar pattern in Jeopardy) and you get your answer.
Anyway, my point is that it doesn't take a huge amount of AI to sift through the first 3 results of Google and try to figure out the response based on a bit of pattern matching, and known questions patterns.
So my point was, that although this is marketed as a huge AI achievement, I am not convinced (although it might just be a poor proof thereof). keyword matching is not AI, and even Google never marketed it as such. This whole "natural language" hype can be a simple "let's ignore the non-useful-too-popular words and concentrate on keywords", which Google does pretty nicely.
I just went to the questions archive (for example http://www.j-archive.com/showgame.php?game_id=3577), and simply pasted to Google the answers that Watson got right. Guess what? if you scratch the "Jeopardy archive" results, which haven't been there while the show was running, you get the correct response in the first or second answers (questions?) from Google. Take that, add a tiny bit of code to clean it up, and you're done.
Not sure if it makes Google more impressive, or IBM less of, but it definitely reduces the AI part to a simple keyword search on the Internet...
Yeah.. it happens a lot in this industry. What I usually do to get a quick introduction about something is to search for it in Wikipedia. They usually have a pretty good description of it in the first paragraph.
Slashdot Mirror
I agree I oversimplified. In the heat of the moment I was amazed by the fact that I checked one after another of the questions, and the response was right there in the title of the first result.
It shouldn't take 4 years and millions of dollars to get from that to winning jeopardy (given also the ~7 seconds of advantage Watson get by getting the answer in the second it shows, while humans must listen to the host and read the question (I can't read it faster when he is reading it out loud.. Maybe Ken can..)).
> That's the tricky part, understanding what is being asked for.
That's right. Don't forget that Watson was programmed for Jeopardy, not generic natural language understanding.
I am pretty sure that by analyzing past questions Watson has so many patterns hard coded in it, so it can quickly devise what kind of information it is looking for, and cross it with a few other search results.
Right, but I just wanted to show that 80% of Watson functionality already exists in Google, so if Watson didn't get it right, it is irrelevant.
> try typing a Jeopardy question into Google sometime.
Have you tried it? Most questions that Watson got right have the result in the title of one of the first three results (don't forget to add -jeopardy, to remove the noise from jeopardy archiving sites.
You're example is a hard one, but most others are not so hard.
Here are a few:
"Elected every 5 years, it has 736 members from 7 parties"
First result - "The European Parliament" - take that verbatim and you win.
"While Maltese borrows many words from Italian, it developed from a dialect of this Semitic language".
First result - "Arabic Language".
"Gambler Charles Wells is believed to have inspired the song "The Man Who" did this "At Monte Carlo"
Second result - "The Man Who Broke the Bank at Monte Carlo (song)". Cross that with the "this" in the question (familiar pattern in Jeopardy) and you get your answer.
Anyway, my point is that it doesn't take a huge amount of AI to sift through the first 3 results of Google and try to figure out the response based on a bit of pattern matching, and known questions patterns.
So my point was, that although this is marketed as a huge AI achievement, I am not convinced (although it might just be a poor proof thereof). keyword matching is not AI, and even Google never marketed it as such. This whole "natural language" hype can be a simple "let's ignore the non-useful-too-popular words and concentrate on keywords", which Google does pretty nicely.
Watson didn't get that answer right, so it is not a good example
I just went to the questions archive (for example http://www.j-archive.com/showgame.php?game_id=3577), and simply pasted to Google the answers that Watson got right. Guess what? if you scratch the "Jeopardy archive" results, which haven't been there while the show was running, you get the correct response in the first or second answers (questions?) from Google. Take that, add a tiny bit of code to clean it up, and you're done.
Not sure if it makes Google more impressive, or IBM less of, but it definitely reduces the AI part to a simple keyword search on the Internet...
Yeah.. it happens a lot in this industry. What I usually do to get a quick introduction about something is to search for it in Wikipedia. They usually have a pretty good description of it in the first paragraph.