| --- |
| task_categories: |
| - text-classification |
| - token-classification |
| language: |
| - en |
| --- |
| |
| # tl;dr: |
|
|
| This is a dataset largely based on CleanCoNLL with some augmentations. |
|
|
| # Details: |
|
|
| ## Base: |
| We started with the CoNLL-2003 dataset, a standard NER benchmark containing English and German text annotated with four entity types: person, location, organization, and miscellaneous. For our evaluation, we focused solely on examples containing the ORG (organization) entity, as these are most relevant to competitor detection. |
| We then applied corrections from CleanCoNLL, a 2023 revision by Rücker and Akbik that addresses annotation errors in the original CoNLL-2003. CleanCoNLL corrects 7.0% of labels in the English dataset, adds entity linking annotations, and maintains the original four entity types. This improved dataset enables more accurate evaluation, with top NER models achieving F1-scores up to 97.1%. |
|
|
| ## Augmentations: |
|
|
| We created two augmented datasets to test specific aspects of competitor detection: |
|
|
| Positive Dataset (with typographical errors): |
| We selected random examples and used the ORG entity as the "competitor" to be detected. We introduced typographical errors to the competitor names by: |
| - Omission: Removing a letter |
| - Transposition: Swapping two adjacent letters |
| - Substitution: Swapping a letter with one found nearby on a US ANSI keyboard layout |
| - Duplication: Selecting a character at random and doubling it |
|
|
| This dataset tests the guardrail's ability to detect variations of competitor names, which is particularly relevant as our solution does not implement fuzzy matching. |
|
|
| Negative Dataset (with distractors) |
| For the negative dataset, we used the original examples containing ORG entities but created a list of "competitors" by randomly selecting companies from the Fortune 500 index (2024), excluding the actual ORG entity in the text. We set the 'has_competitor' flag to 'false' for all examples in this dataset. This evaluates the guardrail's precision in avoiding false positives when no actual competitors are mentioned. |
| |
| # Citations |
| |
| ``` |
| @inproceedings{rucker-akbik-2023-cleanconll, |
| title = "{C}lean{C}o{NLL}: A Nearly Noise-Free Named Entity Recognition Dataset", |
| author = {R{\"u}cker, Susanna and Akbik, Alan}, |
| editor = "Bouamor, Houda and Pino, Juan and Bali, Kalika", |
| booktitle = "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing", |
| month = dec, |
| year = "2023", |
| address = "Singapore", |
| publisher = "Association for Computational Linguistics", |
| url = "https://aclanthology.org/2023.emnlp-main.533", |
| doi = "10.18653/v1/2023.emnlp-main.533", |
| pages = "8628--8645", |
| } |
| |
| @misc{rücker2023cleanconll, |
| title={{C}lean{C}o{NLL}: A Nearly Noise-Free Named Entity Recognition Dataset}, |
| author={Susanna R{\"u}cker and Alan Akbik}, |
| year={2023}, |
| eprint={2310.16225}, |
| archivePrefix={arXiv}, |
| primaryClass={cs.CL} |
| } |
| ``` |
