Francis suggests formatting lex names in DMRS as e.g. verb subscript:v_1(e). The DMRS and long names, when a sentence gets long, can get unreadable and this would help shorten it. Emily disagrees, and finds the subscripts to be more unreadable, and the last bits to be important information.

Emily points out that you're going to keep running into horizontal space problems, no matter how short the predicate names. The real problem is that the DMRS can't break across two lines, which seems like a solvable problem.

Emily: There is some pressure to keep the lexical predicates in roughly the same order as the surface forms due to CS/NLP engineers who are used to using dependency graphs that use the surface forms.

Emily: One thing to consider (with x vs. e in arg0) is prefixing the information to the lexical predicate.

Zhen Zhen: I have a constraint checking for binary feature that must be + that sometimes unifies with underspecified. Would na-or-- or na-or-+ types help with this?

Emily: Here's the hierarchy:

bool na-or-- na-or-+

(The middle row all inherits from luk, + inherits from bool and na-or-+, na inherits from both na types, - inherits from bool and na-or--.)

If you have a type that specifies +, but you don't have a type that specifies -, it's not doing any work for you.

Zhen Zhen: So I need to go through everything and change it?

Glenn and Emily: Yes :( :( exactly :( :(

Zhen Zhen: I'm just starting to use punctuation to control situations with commas.

Emily: Look at the feature R-PERIPH. I'm not saying to use it, but what you're describing is an edge feature. We have the type hierarchy, so you don't have to put it on everything. Use an addendum for the matrix type (unary rule or binary phrase) and add it. The feature basically says "I'm at the edge of a constituent, and you can't modify across this line." Also, Dan has (maybe in ... see saw?) some machinery that allows you to define boundaries before parsing. It's a strategy for dealing with really long sentences -- splitting it into known constituents. What I'm suggesting here is that you treat existing commas in the input as GML. Look at how Dan does that.

Zhen Zhen asks a question about noun features (time, loc, etc)

Glenn suggests that if binary noun features start to get out of hand, especially if you need to exclude one when another is present, you should use a type hierarchy.

Emily: (draws a type hierarchy)

non-loc non-lambda non-place non-time

(everything inherits from ntype. "time" from the first three and "loc" inherits from the last three in the middle row. "other" inherits from all.)

Emily: There's a distinction between saying that only certain nouns can be used adverbially and saying that only certain nouns can be used as vocatives. It's a slippery slope. There are languages that grammaticize animacy. I don't think Chinese is one of them. You can put this in now, but it is not a scalable solution. It will cause some pain later. You can also turn the vocative rule off during parsing so you don't get the extra parses.

LADUW20161117 (last edited 2016-11-17 20:06:01 by LaurieDermer)

(The DELPH-IN infrastructure is hosted at the University of Oslo)