Event-argument relation as a concept

[Guo, Quan]

In ACE05, event-argument needs to be written as concept(s) so that it can be predicted.

[Guo, Quan]

Here is my sample implementation. Please see if this make sense.

Consider we already have these concepts:

phrase = Concept(name='phrase')
pair = Concept(name='pair')
pair.has_a(phrase, phrase)

trigger = phrase(name='trigger')
event = Concept(name='Event')
event.has_a(trigger)

We define concepts based on pairs for argument association:

event_argument = pair(name='EventArgument')
event_argument.has_a(event, phrase)
participant_argument = event_argument(name='Participant')
attribute_argument = event_argument(name='Attribute')

and a shortcut function

def involve(event_type, argument_type, **kwargs):
    for role, concepts in kwargs.items():
        if isinstance(concepts, Concept):
            concepts = (concepts,)
        role_argument = argument_type(name=f'{event_type.name}-{role}')
        ifL(role_argument, ('x', 'y'), andL(event_type, 'x', orL(*concepts, 'y')))

Then we can define concepts and arguments like this:

life = event(name='Life')
be_born = life(name='Be-Born')
involve(be_born, participant_argument, Person=PER)
involve(be_born, attribute_argument, Time=timex2, Place=(GPE, LOC, FAC))

The last two lines are expected to be equivalent to:

be_born_Person = participant_argument('Be-Born-Person')
ifL(be_born_Person, ('x', 'y'), andL(be_born, 'x', orL(PER, 'y')))

be_born_Time = attribute_argument('Be-Born-Time')
ifL(be_born_Time, ('x', 'y'), andL(be_born, 'x', orL(timex2, 'y')))
be_born_Place = attribute_argument('Be-Born-Place')
ifL(be_born_Place, ('x', 'y'), andL(be_born, 'x', orL(GPE, LOC, FAC, 'y')))

When modeling, we might have

pair[be_born_Person] = Some_Learner()
pair[be_born_Time] = Some_Learner2()
pair[be_born_Place] = Some_Learner3()

[Guo, Quan]

@auszok can you review this and see if this is good with the constraint language? Or do you need a running example (maybe with dummy sensor and learners) to test?

[auszok]

The method does not need orL if it only has single concept. Also single variable like 'x' names still need to be in tuples - ('x',).

Corrected below:

def involve(event_type, argument_type, **kwargs):
    for role, concepts in kwargs.items():
        role_argument = argument_type(name=f'{event_type.name}-{role}')
        ifL(role_argument, ('x', 'y'), andL(event_type , ('x',) , concepts, ('y',))))

Similar correction to constraints below:

ifL(be_born_Person, ('x', 'y'), andL(be_born, ('x',) , PER, ('y',) )))

ifL(be_born_Time, ('x', 'y'), andL(be_born, ('x',), timex2, ('y',))))

ifL(be_born_Place, ('x', 'y'), andL(be_born, ('x',), orL(GPE, LOC, FAC, ('y',))))
'''

[auszok]

It would be good to have a running example (maybe with dummy sensor and learners) to test.

[Guo, Quan]

Hi @auszok
Please check out on branch examples/ace-event, folder examples/ACE05, the main_prototype.py. It is a running example.

When I run it, there is some problem in inference with the orL statement in

ifL(be_born_Place, ('x', 'y'), andL(be_born, ('x',), orL(GPE, LOC, FAC, ('y',))))

This is mentioned in a new issue #186