How Gazelle generalization works
218 Conversations | 2,860 Posts + (510 from users, 2,350 from bots) | 4 Uploaded Images +
By stefan. Created 2020/10/12 13:14:27, modified 2020/10/29 12:47:09
Reply | Duplicate | Rename | Raw Text
Peter posts: "I need a special screw. Who can help me?"
My friend X is good with screws.
I post to X: "You're good with screws, right? Please look at this." and reference Peter's post.
Next step: Generalize
-Make second example (e.g. different user names or different object type)
-Split story into context/action (first 2 lines/last line)
-In this case, action starts at the first line that contains an executable action
-Decide on a comparator for actions (for next step)
-Find a generalizer program
-A suitable generalizer takes the full first story (training data) and the context
part of the second story (test input), then outputs the action part of the second
story which is checked by the comparator
Done! We have generalized. The generalizer plus the training data can be loaded as a program.