Access NEL prediction scores across KB candidates
- Dominant language
- Python
- Stars
- 33.9k
- Forks
- 4.7k
- Avg merge
- 3m
- Merged PRs (30d)
- 1
Description
At the time of writing it is not possible to seamlessly access the prediction scores for an Entity Linking model (spaCy 3.4). This impacts the ability to conduct misclassification analyses, improve model results or implement user specific thresholds/heuristics.
Ideally, all scores should be stored and available across viable KB candidates for any one mention (rather than e.g. just the top score) for at least the previous/current epochs, as well as during prediction with a trained model. This should also be compatible with bespoke candidate generation functions.
If the suggestion is implemented it might be useful to include some config parameters to turn the feature off (e.g. to avoid memory restraints). It would also be helpful if the documentation was updated, and if the tutorial code ([nel notebook](https://github.com/explosion/projects/blob/v3/tutorials/nel_emerson/notebooks/notebook_video.ipynb)) was extended to include this or other major updates to the NEL pipeline.
Thank you
Contributor guide
Assessment
This issue has not been assessed yet.