future ideas

[cbergey]

• [ ] repertoire of frames growing vs. repertoire of utterances (# frames/#utts vs. #type/tokens of utts)

• [ ] splitting utterances by sentence into more frames (then re-parse, etc)

• [ ] tf-idf on frames where docs = ages -- which frames are most distinctive of ages?

• [ ] clustering over frame distributions in conversations

• [ ] plain MM

• [ ] sequences of latent states in HMM (n-grams?)

• [ ] look at traces of recursion in viterbi reconstruction vs. free HMM latent state generation

[sdedeo]

I like these!

1. repertoire of frames growing vs. repertoire of utterances (# frames/#utts vs. #type/tokens of utts)

This seems like a really simple thing to do that could tell us a lot; "what's the diversity of slots getting filled in?" Does growth in frames preceed or follow the growth in lexicality? One could also ask questions about the diversity of how a frame is filled in, for example, consider:

"hand me the NOUN"

How many different nouns fill in the NOUN slot? Or even, what's the entropy—call this the "frame-level entropy". How does frame entropy evolve over time? Are new frames low entropy, and get higher? etc. This could help get at the learning question.

3. tf-idf on frames where docs = ages -- which frames are most distinctive of ages?

I am information theory ghost and am here to say that one should do something like partial-JSD! But it's not a big deal.

6. sequences of latent states in HMM (n-grams?)

I like this a lot!

7. look at traces of recursion in viterbi reconstruction vs. free HMM latent state generation

Hmmmmm... not sure what this means, but I like it!