configs:
- config_name: defaultnern
data_files:
- split: bc2gm
path: bc2gm/*.json
- split: bc5chem
path: bc5chem/*.json
- split: bc5disease
path: bc5disease/*.json
- split: chemprotchem
path: chemprotchem/*.json
- split: chemprotgene
path: chemprotgene/*.json
task_categories:
- token-classification
language:
- en
tags:
- NER
This repository contains the data used in the PICLe paper. Is consists of the original and corrupted samples for the following datasets:
- BC2GM
- BC5CHEM
- BC5DISEASE
- CHEMPROTCHEM
- CHEMPROTGENE
The corruption schemes include:
Random ID Label: We replace ground-truth entity labels with random in-distribution entities. For each input sentence, every entity in the ground-truth annotation is replaced by an indistribution (ID) entity randomly sampled from all labels in training samples of the dataset.
Swapped ID Labels: We swap entity labels in the ground-truth demonstrations with the entity labels of a randomly chosen sample in the training split. Contrary to Random ID Label where the number of entities is preserved, the number of entities in each annotation changes compared to the original ground-truth.
Random OOD Label: We replace entity labels in the ground-truth demonstrations with a random out-of-distribution (OOD) English word.
Random OOD Label from Text: We replace ground-truth entity labels with words randomly selected from the sample’s text that are not included in the ground-truth annotation (i.e., not a target entity).
Corrupted OOD Text: We replace the entity mentions in the text with random OOD English words.
Corrupted OOD Text and Label: Similar to Corrupted OOD Text, but we replace ground-truth labels as well, such that the entities in the text and label match.
Corrupted and Shuffled OOD Text, Corrupted and Shuffled OOD Text and Label: Same as their non-shuffled counterpart, but with randomly shuffling the words of the sentence.