Following the example of entity updation in the post on dialogue veneering, here we will try explaining away apparent errors in dialogues by conjuring third parties which can observe and intervene in the process. These parties are not necessarily separate identities but abstractions that help put context in the final output of a veneer which reduces the burden of accurately modeling the language understanding units of the original parties.

The veneering procedure now is something that takes an incoherent dialogue (or just sequence of parses) and returns a coherent polylogue1 with the same parses. Here is an extension of the example from last post2:

Input: A buggy dialogue

Output: A more reasonable trialogue

Influences follow the image below. Note that the dotted edges are not important for the above example.


We can also try to explain the error by personifying the actual issue like this:

Output: A little less appealing alternative

The aim is to create more interesting and meaningful isoparse dialogues. This is basically asking for dialogue generators conditioned on a given list of parses from the two parties while allowing putting in extra context using interventional third parties. Should be easy to do for structured agenda based dialogues, even with bugs since those will be of certain finite kinds.

The harder part will involve creating systems that generalize on more pathological sequences of parses as they will need most of what we call creativity to pull out a good story3. A random example:

This is going towards ideas in computational creativity and thus comes with things like lack of good objective metrics which makes things a little vague. At the same time, our focus on the more abstruse situations like the above makes this interesting.



Abuse of nomenclature mostly.


We probably don't need an explicit intervention for this but it does help.


I still haven't got around to doing this by hand so can't say more at the moment.