[developers] Sweaglesw ERG binaries don't include SEM-I?

Stephan Oepen oe at ifi.uio.no
Tue Mar 14 23:30:39 CET 2017


> Probably the most used is  def_udef_a_q_rel, but I also use unspec_loc_rel,
> poss_rel and can_able rel.

please take a look at the 1214 ‘etc/erg.smi’.  i believe all top-level
abstractions in the list above were preserved (possibly under new
names: ‘unspec_loc’, ‘nn’, and ‘can_able’).  but some of the
sub-hierarchies, e.g. below ‘unspec_loc’ are no longer exposed; they
appeared all too difficult to explain (i.e. document) or interpret
semantically—at least to dan, emily, and myself last year.

when designing the predicate hierarchy for quantifiers in the external
interface, we aimed for simplicity and generalizations that we might
hope to explain to a semanticist.  e.g. the new ‘existential_q’
subsumes what i believe was activated by the old ‘def_udef_a_q’, plus
a few additional quantifiers.

it is an interesting question to try and decide which generalizations
to provide in the external interface, and how to encode relevant
distinctions.  emily, for example, has long argued that dimensions
like definiteness and proximity should be moved to variable
properties, which would reduce the inventory of distinct quantifier
predicates and make strong predictions about which ‘natural’ groupings
one can obtain via underspecification.  i am sympathetic to that
perspective, and i believe we had plans to experiment in this
direction in the ERG ‘trunk’.

much more practically, a key goal in the new SEM-I set-up was better
modularity: a user should be able to take advantage of the added
flexibility in 1214 and augment its SEM-I according to what they need.
in other words, the default SEM-I should be easy to explain and by and
large correspond to distinctions that are grammaticized and ‘make
sense’ semantically.  but it should also be easy to further refine the
SEM-I, e.g. link predicates to an external hierarchy, be it WordNet or
what is useful in a transfer-based MT set-up (for a specific language
pair or groups of languages, maybe).

i just played around with adding the old ‘def_udef_a_q’ on top of the
stock 1214 SEM-I: dropping the file attached below into ‘etc/’ and
adding the following to the end of ‘erg.smi’:

  include: jaen.smi

that much appears to be enough to make the LKB generate the variation
i believe you want:

MRS(172): (lkb::generate-from-mrs (read-mrs-from-file "/tmp/transfer.debug.oe"))
("A dog barks." "The dogs bark." "The dog barks." "Dogs bark.")

—you were maybe the most immediate beneficiary we had in mind when
introducing this new flexibility, i.e. to a large degree de-coupling
the hierarchy of semantic predicates from grammar internals.  in other
words, the interface to generation is now highly configurable by the
user.  the above SEM-I extension should work equivalently in ACE, i
expect; could someone easily confirm that?

best wishes, oe
-------------- next part --------------
A non-text attachment was scrubbed...
Name: jaen.smi
Type: application/smil
Size: 182 bytes
Desc: not available
URL: <http://lists.delph-in.net/archives/developers/attachments/20170314/76c1e76e/attachment.smi>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: transfer.debug.oe
Type: application/octet-stream
Size: 482 bytes
Desc: not available
URL: <http://lists.delph-in.net/archives/developers/attachments/20170314/76c1e76e/attachment.obj>


More information about the developers mailing list