[developers] More SEM-I enhancements

goodman.m.w at gmail.com goodman.m.w at gmail.com
Wed Apr 10 11:26:36 CEST 2019

Hi developers,

The 2016 enhancement to the semantic interface (SEM-I) was a big
improvement over the previous iteration, and I have made use of it for my
MT dissertation work and for Indexed-MRS support in PyDelphin. Thanks again
for all the efforts put into it.

I would like to discuss further enhancements to the SEM-I. I have outlined
these proposals here: http://moin.delph-in.net/SemiRfc#Proposals

(1)--(4) are proposals that were raised previously. Of these, I would most
like to see (3) resolved. Specifically, I would like to stop the dual-use
of 'i' variables as unexpressed 'x' variables as well as the supertype of
'e' and 'x'. I think the following illustrates this:

    _blue_a_1 : ARG0 i, ARG1 i.

The first 'i' is underspecified ("blue [e] car" vs "blue [x] is a color")
and the second 'i' is optionally unexpressed. Relatedly, it was suggested
that the use of square brackets for optionality in the current SEM-I (e.g.,
"ARG0 i, [ ARG1 i ]") should be ignored.

(5) is a new proposal I added as I am thinking about semantic composition.
You could think of this as moving some information from the MRS
fingerprints of the Erg Semantics wikis into the SEM-I. As an example,
consider the ERG's SEM-I entries for _try_v_1 (comments added):

    _try_v_1 : ARG0 e, ARG1 i, ARG2 i.     ; Kim tried an apple.
    _try_v_1 : ARG0 e, ARG1 i, ARG2 h.    ; Kim tried to sleep.

Could we change the latter to capture the subgraph constraints of "try"'s
subject-control? For instance:

    _try_v_1 : ARG0 e, ARG1 i:INDEX #1, ARG2 h:XARG #1.

That is, the XARG of the fragment at _try_v_1's ARG2 is linked up with the
INDEX of its ARG1 (certainly there are better ways of encoding this; its
just for illustration). For comparison, this is what the ErgSemantics wiki

    [ARG0 e1, ARG1+ x2, ARG1+ h3]
    h4:[ARG0 e5, ARG1+ x2]
    { h3 =q h4 }

-Michael Wayne Goodman
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.delph-in.net/archives/developers/attachments/20190410/7a8fe7ee/attachment.html>

More information about the developers mailing list