projects submissions without data/rel_component get_instances modification

[MatthiasMurray]

I want to know if a change I made is relevant enough to be submitted as a pull request on the projects repo, or whether it's best practice to just submit the request and let it be sorted by that process. I'm unsure whether it makes sense since the assets there only refer to one entity type and multiple possible relations between them, and I am not able to open source my own data set that makes the change relevant.

The reason I implemented this is that I have multiple entities, let's call them ENT1, ENT2, ENT3, and I have one relation rel1 that can only be between (ENT1,ENT2) and another relation rel2 that can only be between (ENT3,ENT1) and (ENT3,ENT2). By separating them during training I believe I got a significant boost in performance, but perhaps more importantly I wanted to put the results of the two relation models back together ({**rel1,**rel2}) in a final inference and when I hardcoded the pair types to be considered directly into rel_model.py, it meant they required entirely separate sets of scripts.custom_functions to avoid overlapping/ambiguous methods (very messy). Rather than renaming everything, I found adding the below code to rel_model.py worked better:

@spacy.registry.misc("rel_instance_generator.v2")
def create_instances(max_length: int, entpairs: str) -> Callable[[Doc], List[Tuple[Span, Span]]]:
    def get_instances(doc: Doc) -> List[Tuple[Span, Span]]:
        insttypes = [tuple(x.split('/')) for x in entpairs.split(',')]
        instances = []
        for ent1 in doc.ents:
            for ent2 in doc.ents:
                if ent1 != ent2 and (ent1.label_,ent2.label_) in insttypes:
                    if max_length and abs(ent2.start - ent1.start) <= max_length:
                        instances.append((ent1, ent2))
        return instances

    return get_instances

Then, for the model that could only consider [(ENT3,ENT1),(ENT3,ENT2)], I modified the config like so:

[components.relation_extractor.model.create_instance_tensor.get_instances]
@misc = "rel_instance_generator.v2"
max_length = 100
entpairs = "ENT3/ENT1,ENT3/ENT2"

[polm]

Sorry for the late response to this.

I think this is a great example of one way to handle a particular constraint. Since it's not directly relevant to our sample data though it might seem a little out of place in the sample projects. Since the amount of code involved is rather small, posting it in a Discussion here as a demonstration is probably enough to help people find it if they need it.

[MatthiasMurray]

Thank you! I figured it made more sense in discussions, this is helpful for future thinking about demonstration/presentation vs. contribution.