language:
- en
license: apache-2.0
datasets:
- CoNLL-2003
metrics:
- F1
This is a T5 small model finetuned on CoNLL-2003 dataset for named entity recognition (NER).
Example Input and Output: “Recognize all the named entities in this sequence (replace named entities with one of [PER], [ORG], [LOC], [MISC]): When Alice visited New York” → “When PER visited LOC LOC"
Evaluation Result:
% of match (for comparison with ExT5: https://arxiv.org/pdf/2111.10952.pdf):
| Model | ExT5_{Base} | This Model | T5_NER_CONLL_OUTPUTLIST |
|---|---|---|---|
| % of Complete Match | 86.53 | 79.03 | TBA |
There are some outputs (212/3453 or 6.14% that does not have the same length as the input)
F1 score on testing set of those with matching length :
| Model | This Model | T5_NER_CONLL_OUTPUTLIST | BERTbase |
|---|---|---|---|
| F1 | 0.8901 | 0.8691 | 0.9240 |
**Caveat: The testing set of these aren't the same, due to matching length issue... T5_NER_CONLL_OUTPUTLIST only has 27/3453 missing length (only 0.78%); The BERT number is directly from their paper (https://arxiv.org/pdf/1810.04805.pdf)