Closed lichtefeld closed 4 years ago
@denizbeser Before I go to implement this solution does it make sense to you?
@LichMaster98 This makes sense to me. Let me know if you run into any issues!
BEFORE
and the destination is marked as AFTER
.It may make sense for whoever is going to do this to make notes of any paths they are unsure how to implement and get feedback from
Using this comment to track things to discuss/uncertainties:
For FALL
, we have an issue with requiring distance or direction in our region since in our curriculum, we currently can either fall from DISTAL
or PROXIMAL
but must specify one in the ontology, which leads to an orphaned node corresponding to whichever one we've picked since it's always in the semantics. Example: slot1_falls_toward_slot2.pdf has an orphaned DISTAL
since I defined FALL
to have a path from DISTAL
to the ground towards the ground.
For GIVE
since we don't allow two edges between two nodes, if we express relative size between the subject and direct object, this edge label over-rides the edge label indicating that the subject possesses the object at the beginning of the relation. I disabled this relative size to preserve the possession since my judgement was that possession is more salient for the giving action than relative size but I am happy to change it if others disagree.
A lot of over-specification happens with the flying curriculum since we only have one object that can fly but perhaps this is alright. Currently, there is always a path for flight and whether it is towards or away from the ground is specified with flying up or down. For flying over and under, we always assume DISTAL
-- I'm not sure if this is an assumption we wish to make or not.
Should DRINK_CONTAINER_AUX
have contacts with the DRINKER
? I don't know that this is a necessary specification but I think it is true for most drinking instances. For DRINK
, we need to check DRINK FROM
with a native speaker for Chinese.
JUMPING
-- should our origin and goal both be the ground? I tend to think so but I'm not sure.
The learned semantics for SIT
and SIT DOWN
are the exact same, as are those for SIT ON
and SIT DOWN ON
For some reason with transitive moving, we can only have a path towards MOVE_GOAL
for either the agent or theme, not both. It's currently set to be the theme so I left it as that.
Since our GO
semantics are defined such that the object is always proximal to the goal at the end of the action, it doesn't make sense to go to under and DISTAL
from the object and the semantics break if we try to so I have disabled that. I have also removed the DISTAL/PROXIMAL
alternation from verbs with dynamic prepositions as this causes the same issue of a broken path.
COME
currently only expresses the relative location of the speaker and goal for COME DOWN
because elsewhere we end up combining coming to speaker and coming to object semantics.
Our BESIDE GOAL
semantics are currently way over-generalized because we try both left and right and this cuts off the edge between proximality and the goal, for example: slot1_pushes_slot2_beside_slot3.0.pdf
In general, we can't alternate things like DISTAL/PROXIMAL
or direction in a path for the same situation without just breaking it for verbs with dynamic prepositions.
Currently our
SpatialPath
is made up of anoperator
and areference_object
(or region). To solve disambiguation between the source and destination objects theSpatialPath
should just be refactored to have tworeference_objects
asource_object
anddestination_object
This will require going through all our situation dictionaries and templates to update for the new structure of the
SpatialPath
which is probably a several hour process.