Unintended machine learning biases as social barriers for persons with disabilitiess

Abstract

Persons with disabilities face many barriers to full participation in society, and the rapid advancement of technology has the potential to create ever more. Building equitable and inclusive technologies for people with disabilities demands paying attention to more than accessibility, but also to how social attitudes towards disability are represented within technology. Representations perpetuated by machine learning (ML) models often inadvertently encode undesirable social biases from the data on which they are trained. This can result, for example, in text classification models producing very different predictions for I am a person with mental illness , and I am a tall person . In this paper, we present evidence of such biases in existing ML models, and in data used for model development. First, we demonstrate that a machine-learned model to moderate conversations classifies texts which mention disability as more "toxic". Similarly, a machine-learned sentiment analysis model rates texts which mention disability as more negative. Second, we demonstrate that neural text representation models that are critical to many ML applications can also contain undesirable biases towards mentions of disabilities. Third, we show that the data used to develop such models reflects topical biases in social discourse which may explain such biases in the models - for instance, gun violence, homelessness, and drug addiction are over-represented in discussions about mental illness.

Algorithmic Fairness Datasets: the Story so Far

“…Good documentation should discuss and explain features, providing context about who collected and annotated the data, how and for which purpose (Gebru et al, 2018;Denton et al, 2020). This provides dataset users with information they can leverage to select appropriate datasets for their tasks and avoid unintentional misuse (Gebru et al, 2018).…”

Children in the digital environment

“…Violent and pornographic content can cause children shock and disgust. A study from 2020 revealed that in the European Union the most reported harmful content that children were exposed to (at least monthly) were hate messages 1 (average of 17%) followed by violent images (average of 13%) (Smahel et al, 2020 [14]) (Council of Europe, 2018 [15]) This report also found that exposure to different kinds of harmful content is interrelated. For instance, if a child sees one type of harmful content, it is more likely that the same child will also report seeing other types of harmful content (Smahel et al, 2020 [14]).…”

A Systematic Review of Machine Learning Algorithms in Cyberbullying Detection: Future Directions and Challenges

“…Few works are found to solve this problem, and more research is required to produce the cyberbullying detectors using machine learning algorithms that are unbiased and transparent. Several works proposed methods to mitigate the unintended bias in the word embeddings [171][172][173]. Similarly, measuring and mitigating the classification algorithms' bias is also important for a fairer performance of the classifier [173].…”

“…Several works proposed methods to mitigate the unintended bias in the word embeddings [171][172][173]. Similarly, measuring and mitigating the classification algorithms' bias is also important for a fairer performance of the classifier [173]. Data is collected and annotated by subjective human annotations.…”