Computer system automatically solves word problems

newsoffice.mit.edu

40 points·ColinWright··8 comments

Paper located:

http://people.csail.mit.edu/nkushman/papers/acl2014.pdf

> On a certain day, 278 people entered the park.

I didn’t test this with robots, but I tested this with humans (first year university students). It’s more difficult than expected, because it has to be translated as 1 * X + 1 * Y = 278, but neither of the “1” coefficients are written in the text. It breaks the “easy” algorithm that is “copy all the numbers in the matrix, and hope that the solution is related to the problem”.

Peter Norvig covered STUDENT (http://en.wikipedia.org/wiki/STUDENT_(computer_program)), a program that solves algebra word problems, in "Paradigms of Artificial Intelligence Programming: Case Studies in Common Lisp" back in 1992.

Perhaps also of interest in this context is the work of Gordon Novak (http://www.cs.utexas.edu/users/novak/) in the 1970s, related to AI/NLP interpretation of elementary physics problems [1].

[1] http://www.cs.utexas.edu/users/novak/ijcai77.html

Yes but it doesn't understand. Let's say I gave it something more complex like, "let's call plus one meow. What is meow of two?" I bet most third graders would get this correct.

Why is that relevant? Anyway, computers are already good at syntax macro expansion. I feel like that's even less "understanding."

The way you worded it gave me a big pause (non native speaker). Had to read it 5 times.

No link to the paper?

Reply on news.ycombinator.com