In this article, we present new empirical evidence to demonstrate the severe limitations of existing machine learning content moderation methods to keep pace with, let alone stay ahead of, hateful language online. Building on the collaborative coding project “AI4Digntiy,” we outline the ambiguities and complexities of annotating problematic text in AI-assisted moderation systems. We diagnose the shortcomings of the content moderation and natural language processing approach as emerging from a broader epistemological trapping wrapped in the liberal-modern idea of “the human.” Presenting a decolonial critique of the “human vs machine” conundrum and drawing attention to the structuring effects of coloniality on extreme speech, we propose “ethical scaling” to highlight moderation process as political praxis. As a normative framework for platform governance, ethical scaling calls for a transparent, reflexive, and replicable process of iteration for content moderation with community participation and global parity, which should evolve in conjunction with addressing algorithmic amplification of divisive content and resource allocation for content moderation.

The size of the vocabulary is a central design choice in large pretrained language models, with respect to both performance and memory requirements. Typically, subword tokenization algorithms such as byte pair encoding and WordPiece are used. In this work, we investigate the compatibility of tokenizations for multilingual static and contextualized embedding spaces and propose a measure that reflects the compatibility of tokenizations across languages. Our goal is to prevent incompatible tokenizations, e.g., "wine" (word-level) in English vs. "v i n" (character-level) in French, which make it hard to learn good multilingual semantic representations. We show that our compatibility measure allows the system designer to create vocabularies across languages that are compatible -a desideratum that so far has been neglected in multilingual models. * Equal contribution.

Warning: This work contains strong and offensive language, sometimes uncensored.To tackle the rising phenomenon of hate speech, efforts have been made towards data curation and analysis. When it comes to analysis of bias, previous work has focused predominantly on race. In our work, we further investigate bias in hate speech datasets along racial, gender and intersectional axes. We identify strong bias against African American English (AAE), masculine and AAE+Masculine tweets, which are annotated as disproportionately more hateful and offensive than from other demographics. We provide evidence that BERT-based models propagate this bias and show that balancing the training data for these protected attributes can lead to fairer models with regards to gender, but not race.

No abstract

Research to tackle hate speech plaguing online media has made strides in providing solutions, analyzing bias and curating data. A challenging problem is ambiguity between hate speech and offensive language, causing low performance both overall and specifically for the hate speech class. It can be argued that misclassifying actual hate speech content as merely offensive can lead to further harm against targeted groups. In our work, we mitigate this potentially harmful phenomenon by proposing an adversarial debiasing method to separate the two classes. We show that our method works for English, Arabic German and Hindi, plus in a multilingual setting, improving performance over baselines.