Yes we have removed rules which have lower probabilities than a certain threshold.

Another way of asking the question is: do you understand why you are able to learn the languages with your algorithm.
In particular, do you actually prune rules with non-zero weights?

I don’t even know whether stochastic indexed grammars have been considered before. I don’t think there would be a direct translation between the two stochastic formalisms. The stochastic part wouldn’t correspond. I imagine the situation would be similar to what happens with, e.g., left-corner transform of CFGs.
(Maybe it’s somewhat easier to work with OI context-free tree grammars, rather than OI macro grammars, as the starting point.)

An interesting question. Do you know whether *stochastic* OI macro grammars are equivalent to *stochastic* indexed grammars?
If the answer is yes, and there is a constructive proof of the equivalence, you can certainly obtain stochastic indexed grammars from our algorithm. Otherwise, I am not sure.

Thank you for asking. I’m afraid I don’t know whether we can get an algorithm in a similar way to the two previous talks. There may be different approaches for learning (non-stochastic) macro grammars and they could be investigated further, but our experiments were intended to confirm whether we can get optimal probability assignments of stochastic macro grammars through our algorithm.