Multiple Subjects in a Paragraph Does Not Work

[jsnhff]

Discovered and written up by @estambolieva (added here by me, jsnhff)

A paragraph starts with a mention of the protagonist. Protagonist coreferences follow. Then another proper name is introduced and when this proper name is of the same gender as the protagonist.

Action to be taken: All protagonist’s coreferences following the same-sex proper name are removed from the protagonist’s coreference cluster as we expect them to be related to the proper name.

Reference Excerpts: Rabbit, Run

This problem had been mentioned in 2019 in the neuralcoref library as well (see here) however the thread was closed without it being resolved.

A paragraph starts with a mention of the protagonist. Protagonist coreferences follow. Then another proper name is introduced and when this proper name is of the same gender as the protagonist.

Action to be taken: All protagonist’s coreferences following the same-sex proper name are removed from the protagonist’s coreference cluster as we expect them to be related to the proper name.

Extra: The protagonist’s name is mentioned again in the same paragraph. Then the coreferences to be removed are only the ones between the proper name and the next mention of the protagonist.

Reference Excerpts: 4th paragraph, Pride & Prejudice

What happens when the mention of the next proper name happens when between quotes.

Action to be taken: ignore all mentions of proper names between quotes. Reference Excerpts: the Sound and Fury

Future: do not interrupt the coreference cluster when the proper name interruption in the middle is for a person from a different gender.

Example (see this Trello ticket):

This is an example from the Sound and Fury. The mention of Jason nullifies the coreferences to the now-regendered-to-Emilio protagonist even when the coreference cluster has identified them correctly.

Sometimes proper names do not interrupt the protagonist’s coreferences such as Hedwig and Harry in this example below. Action to be taken: investigate more and fix. This is very important.