Specs for the anticipated behavior of ICONS in generation, from the point of view of those using them for information structure:
Original set of specs (by Emily and Sanghoun)
The behavior we would like to see is partially documented in the files eng.txt and jpn.txt (in this tarball, along with grammars), which associate input sentences (for parsing) with the realizations the generator currently finds. Among those, we have marked the ones we would like to not see.
In general, the idea is that for any given pair of indices, if there is an ICONS element in the MRS associated with a realization that realization is valid iff:
- There is no ICONS element associated with the same pair of indices in the input; or
- Any ICONS element associated with the same pair of indices in the input is compatible in its type with the ICONS element in the candidate realization.
(I'm not sure as I type this what to do about cases where there are multiple ICONS elements for the same pair of indices. As far as information structure is concerned, we wouldn't expect to have more than one. Of course, if we're putting other stuff into ICONS, then there might be ICONS constraints of different types. All of the above would then be only applicable to subtypes of info-str.)
Initial follow up during ACE development
Emily: I see that the condition you describe could be summarized as "the input ICONS list and the realization's ICONS list are unifiable", under a reasonable extension of the term.
Then Woodley did some magic, and ACE now treats ICONS the way we would like them to be treated. Woodley, can you reconstruct what further decisions you had to make to operationalize the specs above?
Woodley: In answer to your question, what I did in ACE is fairly straightforward. To begin with, I made ICONS processing optional, enabled by the "enable-icons := yes." configuration parameter (and a couple of accompanying settings), which is off by default. When ICONS processing is on, all MRSes (extracted from feature structures, parsed from strings/files, or formatted for output) have an additional ICONS list. It follows the form of the HCONS list in just about every way, except that we expect a richer hierarchy for the constraints types.
The post-generation subsumption test generally verifies that a candidate realization's MRS contains all of the information specified on the input MRS (and possibly more). For HCONS, at least in ACE, that means ensuring that there is at least one unique realized HCONS element that is compatible with each input HCONS element. I used almost exactly the same algorithm to test ICONS compatibility -- i.e. I look for an injective mapping from the input ICONS list to the realization's ICONS list, such that (1) variable identities are compatible, (2) variable properties are subsumed, and (3) the constraint's type is subsumed. If, after satisfying all of the input ICONS elements, there are leftover realization ICONS elements, I don't currently consider that grounds for rejecting the result. This allows inputs with empty ICONS lists to still generate, for instance.