WIKIBIAS: Detecting Multi-Span Subjective Biases in Language

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 sentencelevel bias types and token-level biased segments. We present systematic analyses of our dataset and results achieved by a set of state-ofthe-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. 1

Reinforced Sequence Training based Subjective Bias Correction

“…Encouragingly, previous research has begun to make progress in identifying prejudiced or emotive language used in factual statements (Recasens et al, 2013;Bhosale et al, 2013;Misra and Basak, 2016;Hube and Fetahu, 2018;Zhong et al, 2021;Madanagopal and Caverlee, 2022) and a few studies have begun to investigate subjective bias correction (Pryzant et al, 2019;Liu et al, 2021;Zhong et al, 2021). These approaches, however, typically face a number of key challenges: Noisy and Limited Training Data.…”

“…Training-Testing Mismatch. Second, existing approaches are primarily trained on maximizing the likelihood of each token of the target sequence (Pryzant et al, 2019;Zhong et al, 2021) by relying on the input sequence and previous ground-truth token, but testing is done on the entire input and output sequence. The training is conducted using token-level objective functions and tested using sentence-level evaluation metrics, such as BLEU.…”

“…Since minimizing L CE makes the model generate text close to the human-reference edits, the training data needs to be of high-quality. Since the only large parallel corpus available for bias correction is noisy, such likelihood optimized models have limited performance (Pryzant et al, 2019;Zhong et al, 2021). Additionally, these models do not produce the best results for metrics during evaluation because the training is not optimized to generate text that addresses evaluation metrics (such as BLEU or Content Similarity).…”

“…For training a cross-domain bias classifier, we used the dataset assembled in a prior study (Madanagopal and Caverlee, 2022). For evaluation, we used WIK-IBIAS (2,117 biased and 2,911 neutral), a highquality dataset that is manually curated by (Zhong et al, 2021).…”

State-of-the-art generalisation research in NLP: A taxonomy and review

“…In-context finetuning A different class of studies that considers task evaluation in the prompting setup are those that finetune a pretrained model with prompts from one set of tasks and then evaluate them on another set of tasks (e.g. Zhong et al, 2021;Sanh et al, 2022;Wei et al, 2022). Parallel to the term 'in-context learning', this scenario is often referred to with the term in-context finetuning.…”