Cannot use ConllCorefScores for general coref tasks

[david-waterworth]

Is your feature request related to a problem? Please describe.
The coreference model implement in allennlp coref.py is based on Higher-order Coreference Resolution with Coarse-to-fine Inference by Lee et al., 2018.

This model considers for a span i, all antecedents j, j<i and introduces a dummy cluster to represents spans with no antecedent. In [DWIE: an entity-centric dataset for multi-task document-level information extraction] by Zaporojects et al. 2021 this was seen as a limitation, they allowed singleton clusters j=i and eliminated the dummy cluster. In my current work I've combined both, a span can either be invalid (belong to the dummy cluster), be a singleton or have an antecedent.

The problem I've run into is due to a single assertion in ConllCorefScores the scores cannot be computed when the model is capable of predicting singletons

https://github.com/allenai/allennlp-models/blob/5012f2387decc806152fcba6ad81345b7627fc2a/allennlp_models/coref/metrics/conll_coref_scores.py#L107

I realise the code is correct in that it's based on models which cannot predict singleton clusters, but it would be nice if it were more general-purpose - specifically the only issue being a single assertion that is only valid for specific models.

[AkshitaB]

@david-waterworth Will it suffice to add an attribute to the metric class that describes whether the model produces singletons or not, and use that to check for that assertion?

[david-waterworth]

@AkshitaB yes that would be fine, in my case when there are singletons the assert could check assert i >= predicted_index and without singletons the existing check assert i > predicted_index would be retained. Not sure if there's other use cases.