Leveraging DELPH-IN grammars to develop educational materials
Discussion at the Cambridge Summit led by Michael Goodman; scribed by Kristen Howell (2019).
Participants: Michael Goodman (MG), Emily Bender (EB), Dan Flickinger (DF), Woodley Packard (WP), David Inman (DI), Francis Bond (FB), Guy Emmerson (GE), Anne Copestake (AC)
MG: For some tasks, MRSs must conform to the semantic model to be useful. The model involves several structures, including: variable hierarchy, role inventory, property hierarchy, predicate synopsis. MG: Automatic generation of the SEM-I is not bug-free MG: The wiki includes a list of open issues: http://moin.delph-in.net/SemiRfc
EB: (Re the issue: "Use SEM-I to encode order of roles") Why does order matter? Notionally it doesn't DF: Normalization
Re the issue "Encode HCONS and ICONS..." DF: ICONS are contributed by phrase structure rules, not lexical entries DI: Focus pronouns are topic pronouns WP: Construction based predicates should be in there to DI: Who's consuming the SEM-I? MG: I do. Ace uses it to compile a grammar FB: Some features are only grammar internal. The semi are the features that we think should be visible from the outside
MG: Can we make the model a curated resource? I'm hesitant to put the grammar constraints on this resource, but it would be nice if we could automaticall get at it. WP: What's an example of something you don't think we could get at MG: points to some examples where try is a control verb.
GE: Could you have some construction higher in the tree? EB: Control is best viewed as a lexical property. For optionality, there is a whole other discussion to be had there. OPT is also lexical. But what do we mean by optionality? Doesn't it mean the argument is dropped or that it's not linked to anything in the MRSs AC: Well it doesn't need to be dropped DF: That gets vexed because we pretend that English requires an ARG1. So we never say that ARG1 is optional in the SEM-I WP: It's useful for the consumer of the MRS to know what the pattern is. Eg. with active passive MG: In the semantic model there are no constraints on what can be dropped. We don't know if an argument isn't valid for a predicate or if it's missing AC: Origionally we thought that the SEM-I would help here WP: WHat are the practical uses of the SEM-I other than predicting what could be generated MG: Indexing MRSs with fixed arity, composition, validation EB: WHy composition MG: I use it externally to the grammar to construct an MRS directly AC: So the idea is for it to be useful as an API, so you know what can usefully be fed into a grammar. We've used it for that on and off but it's never been perfected
MG: Between 1214 and 2018, I noticed some differences: In an example with numbers the ARG0 of plus is an e in 1214 and an i in 2018. Why? DF: That was intentional. The plus is more like a predication and I'm moving more and more towards treating them as "e"s in their ARG0s DF: "23 is an odd number" we want that to be an individual. Others can be predications. I suppose I could coerce them into some eventuality so that you never see the i in the resulting MRS (when you have a plus or a times) MG: If the i is meant to be underspecified, that's fine, but it should be documented ...looking at another example with numbers DF: I don't know if I have a good justification for that. I don't know if I should be calling that an eventuality (is it a predicate?), so I took a more neutral view EB: Is predicativeness the test that something should be an eventuality? DF: It's a convincing test. I need an ARG0 to sew together modifiers EB: So "very" won't modify a noun, so it needs an event ARG0? DF: Yes GE: If we have a model structure for evaluating the truth of a sentence, we can leave it underspecified DI: But then in can be instantiated as an individual AC: This is not an easy problem, if we want to return to it, we should give it its own SIG DF: I've dumped this problem. And now Mike sees why
A new variable heirarchy is discussed. Scribe couldn't keep up. Lot's of discussion regarding when/why/how quantification is required.
DF: Alex L. has always scolded us on allowing arguments of different types to have different predicates AC: Anne doesn't really care, we can just invent some more names DF: Should that be part of the external MRS? From where does that arrive? MG: If it's not in the MRS that came out of the grammar, it doesn't really help. I have to choose it its /HEQ or /NEQ AC: There needs to be a deterministic way to decide how to add another argument
DF: Constructive take away: we should change the way we check for quantifiers BC: Caveat, util has to be used for additional well formedness, so we might need to change that too