_dist in Meaning: either make tuple or eliminate

[shanest]

_dist in Meaning is still a dict. I thought about quickly changing it to a tuple, but wanted to discuss another possibility: just eliminating it all together. A Meaning still has a Universe, which itself has a prior, so we can get a distribution for the Meaning from the Universes. Are there situations where we want the Meaning to be a distribution that comes from some other source than this?

[nathimel]

Ah, I tend towards making it a tuple. I had similar thought process when recently working on the recent hashability issues but for whatever reason forgot to implement.

And yeah, while I think letting a Meaning encode the literal distribution over referents for its expression is clearly an important distinction from the prior ( e.g., color names), we haven't had a use case for this yet since the research ultk supports mainly assumes that every referent is equally, for each Expression/Meaning. But it's nice to have since much of the effcomm research does not make this assumption.

[shanest]

Yeah, I totally agree that hard-coding the uniform distribution is the "worst" solution. So I think the options are (i) dist is an arbitrary tuple; if it's not specified in __init__, we could default to initialize it according to self.universe.prior; or (ii) no dist at all in Meaning, always just appealing to the Universe. The question is whether (ii) is also too restrictive 🤔

[nathimel]

Oh, I see. What does it mean to initialize according to self.universe.prior? To take only the subset of probabilities from the universe prior corresponding to the Meaning.referents, and then renormalize?

[nathimel]

This commit has the tuple implementation, but defaults to the uniform distribution.

N.B.: the current modals implementation and indefinites example results assume the uniform distribution even if that is the 'worst', so maybe I'd add a kwarg for uniform initialization to replicate that behavior if we chose not to uniform initialize meanings by default

[nathimel]

After further consideration we have decided to explore letting _dist be a frozenset of (Referent, float) tuples. See #33