Investigating and Mitigating Biases in Crowdsourced Data

Abstract

It is common practice for machine learning systems to rely on crowdsourced label data for training and evaluation. It is also wellknown that biases present in the label data can induce biases in the trained models. Biases may be introduced by the mechanisms used for deciding what data should/could be labelled or by the mechanisms employed to obtain the labels. Various approaches have been proposed to detect and correct biases once the label dataset has been constructed. However, proactively reducing biases during the data labelling phase and ensuring data fairness could be more economical compared to post-processing bias mitigation approaches. In this workshop, we aim to foster discussion on ongoing research around biases in crowdsourced data and to identify future research directions to detect, quantify and mitigate biases before, during and after the labelling process such that both task requesters and crowd workers can benefit. We will explore how specific crowdsourcing workflows, worker attributes, and work practices contribute to biases in the labelled data; how to quantify and mitigate biases as part of the labelling process; and how such mitigation approaches may impact workers and the crowdsourcing ecosystem. The outcome of the workshop will include a collaborative publication of a research agenda to improve or develop novel methods relating to crowdsourcing tools, processes and work practices to address biases in crowdsourced data. We also plan to run a Crowd Bias Challenge prior to the workshop, where participants will be asked to collect labels for a given dataset while minimising potential biases. CCS CONCEPTS• Human-centered computing → Collaborative and social computing.

LINGO : Visually Debiasing Natural Language Instructions to Support Task Diversity

“…Several techniques have been proposed to handle such bias post‐creation, including improving linguistic diversity of samples [YZFBC * 20, LML * 19, SYH20] and augmenting data with adversarial samples intended to fool the model [WRF * 19,KBN * 21,TYLB * ]. Similarly, there is evidence that natural language instructions provided by dataset creators during crowdsourcing influences crowdworkers to follow specific patterns during sample creation [GGB19,PMGB22, HSG * 21]. These patterns, termed as ‘instruction bias,‘ propagate to the dataset and are subsequently over‐represented in the collected data.…”

“…Unfortunately, prior work in crowdsourcing has shown that task instructions provided by dataset creators often influence crowd‐workers to follow specific patterns during instance creation. This leads to collection of biased data, which inflates model performance [GGB19, PMGB22, HSG * 21]. This is particularly critical in high‐risk domains such as healthcare [MA21a], where incorrect answers to a task can prove fatal.…”

How Crowd Worker Factors Influence Subjective Annotations: A Study of Tagging Misogynistic Hate Speech in Tweets

“…Similar results were evident from work by Hube, Fetahu, and Gadiraju (2019), who found that crowd workers with strong opinions produce biased annotations, even evident among experienced workers. Furthermore, demographics, location, worker context, and work en-vironment have been found to impact annotation quality in both subjective and objective tasks (Hettiachchi et al 2021).…”

Don’t Blame the Annotator: Bias Already Starts in the Annotation Instructions

“…Crowdsourcing has been a widely adapted approach to create large scale datasets such as SQUAD 1.1 (Rajpurkar et al, 2016(Rajpurkar et al, , 2018, DROP (Dua et al, 2019), QUOREF and many more (Najafabadi et al, 2015;Callison-Burch and Dredze, 2010;Lasecki et al, 2014;Zheng et al, 2018;Chang et al, 2017). Many past works investigate different types of bias in crowdsourcing datasets such as cognitive bias (Eickhoff, 2018), annotator bias (Gururangan et al, 2018;Geva et al, 2019), sampling bias (Hu et al, 2020), demographic bias (Rahmani and Yang, 2021) and others (Hettiachchi et al, 2021). Many works on bias in NLU benchmarks focus on biases resulting from the crowdsourcing annotations, and how annotator-specific patterns create biases in data (Geva et al, 2019).…”

“…Despite the vast success of this method, past studies have shown that data collected through crowdsourcing often exhibit various biases that lead to overestimation of model performance (Schwartz et al, 2017;Gururangan et al, 2018;Poliak et al, 2018;Tsuchiya, 2018;Mishra et al, 2020a;Mishra and Arunkumar, 2021;Hettiachchi et al, 2021). Such biases are often attributed to annotator-related biases, such as writing style and background knowledge (Gururangan et al, 2018;Geva et al, 2019) (see more discussion on related work in §A).…”