Create a lesson grammar from examples #90
Description
The general idea is to filter the RGL (augmented with a lexicon), so that it accepts some given sentences, and rejects some other given sentences.
This can be viewed as an optimization problem: find a set of grammar rules (subtrees) such that their size (or number) is the smallest. Here is an abstract procedure:
- We start with a very general grammar (in our case the resource grammar of GF), and a number of examples that are accepted by the grammar.
- What we want is to limit/restrict the grammar as much as possible, but so that it still accepts the examples.
- The phrase "limit/restrict as much as possible" is a bit unclear. What we mean is that we want an as small grammar as possible (some kind of Kolmogorov complexity), perhaps also to reduce the number of parse trees as much as possible (or the size of the trees).
- A possible extension could be to also give negative examples, with the meaning that the final grammar should not accept these.
- Another possible variant would be to not just filter the original grammar rules, but to allow merging of rules. (Meaning that "A -> B C" and "B -> C D" can be merged to the rule "A -> C C D")