Skip to content

Evaluation of Textual Event Mention #27

Description

@seebergerph

Hi all,

I noticed that the evaluation script scripts/evaluation/test_joint_object.sh uses the argument --keep_events 1.

This setting only loads M2E2 sentences containing at least one event, which means the majority of test sentences without any events are ignored. When I run the code with --keep_events 1, I obtain better performance in terms of textual event mentions than reported in the paper (P: 45.6, R: 72.0, F1: 55.8).

However, when I switch to --keep_events 0, performance drops significantly due to obvious precision issues (P: 14.4, R: 72.0, F1: 24.0).

What is the intuition behind dropping sentences without ground-truth events and which setup was used for the reported results? I’d be happy to discuss further if anyone can clarify this.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions