WIKIBIAS: Detecting Multi-Span Subjective Biases in Language

This repo contains codes for the following paper:

@inproceedings{zhong-etal-2021-wikibias-detecting,
    title = "{WIKIBIAS}: Detecting Multi-Span Subjective Biases in Language",
    author = "Zhong, Yang  and
      Yang, Jingfeng  and
      Xu, Wei  and
      Yang, Diyi",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2021",
    month = nov,
    year = "2021",
    address = "Punta Cana, Dominican Republic",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.findings-emnlp.155",
    pages = "1799--1814",
    abstract = "Biases continue to be prevalent in modern text and media, especially subjective bias {--} a special type of bias that introduces improper attitudes or presents a statement with the presupposition of truth. To tackle the problem of detecting and further mitigating subjective bias, we introduce a manually annotated parallel corpus WIKIBIAS with more than 4,000 sentence pairs from Wikipedia edits. This corpus contains annotations towards both sentence-level bias types and token-level biased segments. We present systematic analyses of our dataset and results achieved by a set of state-of-the-art baselines in terms of three tasks: bias classification, tagging biased segments, and neutralizing biased text. We find that current models still struggle with detecting multi-span biases despite their reasonable performances, suggesting that our dataset can serve as a useful research benchmark. We also demonstrate that models trained on our dataset can generalize well to multiple domains such as news and political speeches.",
}

If you would like to refer to it, please cite the paper mentioned above.

Prerequisites

python==3.6.8
torch=1.5.0
scikit-learn==0.23.1
transformers==2.8.0
pytorch-transformers=1.2.0
nltk=3.4.5

If you are using the conda enviroment, you can install all dependencies with the provided requirments.txt file.

Data

Train/Dev/Test Data

Please download all ourd data from here and put under data. There should be four folders:

class_binary
class_finegrained
tag_binary
tag_finegrained

Each folder contains a train, dev, and test split. We provide the detailed data format in the README file under data.

Data Annotation

We release the data annotation interfaces and instructions under data/annotations.

Training Models

Classification

Please go to ./src/Classification.

The following command fine-tunes an bert-base sentence classification model for binary bias detection.

python train.py --device [desired cuda id] 
                --batch-size 16 
                --epochs 3 
                --max_len 128 
                --lr 2e-5
                --pretrained-ckpt "ckpt/*saved_ckpt*" [optional]

The following command fine-tunes an bert-base sentence classification model for finegrained bias detection.

python train_finegrained.py --device 0 [desired cuda device] 
                --batch-size 16 
                --epochs 3 
                --max_len 128 
                --lr 2e-5
                --pretrained-ckpt "ckpt/*saved_ckpt*" [optional]

To evaluate the results, run

python eval.py/eval_finegrained.py --device 0 [desired cuda device] --eval_ckpt "ckpt/*saved_ckpt*"

Tagging

Please first change directory to ./src/Tagging.

Training BERT model

Please run ./run_bert_ner.sh to train the BERT baseline model without extra linguistic features.

Training BERT-LING model

Please run ./run_bert_ner_extra_feature.sh to train the BERT baseline model with extra linguistic features.

Training BERT-LING model for fine-grained settings

Please run ./run_bert_ner_extra_feature_fine_grained.sh to train the BERT baseline model with extra linguistic features in the fine-grained setting.

Evaluating models

Please follow the example jupyter notebook (eval.ipynb) for evaluating the model on test set for Exact and Partial F1 scores.

cjer / WIKIBIAS