RECAST: Enabling User Recourse and Interpretability of Toxicity Detection Models with Interactive Visualization

  title={RECAST: Enabling User Recourse and Interpretability of Toxicity Detection Models with Interactive Visualization},
  author={Austin P. Wright and Omar Shaikh and Haekyu Park and Will Epperson and Muhammed Ahmed and Stephane Pinel and Duen Horng Chau and Diyi Yang},
  journal={Proc. ACM Hum. Comput. Interact.},
Fig. 1. The Recast user interface. A. Toxicity score of overall input text shows edits’ effect on toxicity in real time. B. Words whose possible alternatives have strong potential for toxicity reduction are highlighted in yellow. C. Usage guide for Recast’s capabilities. D. Underline opacity visualizes model’s attention on words, including those without alternatives, to inform users about which words contribute important context (e.g., “kid” is underlined, because toxicity towards a kid… 

Figures and Tables from this paper

End-User Audits: A System Empowering Communities to Lead Large-Scale Investigations of Harmful Algorithmic Behavior

In an evaluation of end-user audits on a popular comment toxicity model with 17 non-technical participants, participants both replicated issues that formal audits had previously identified and also raised previously underreported issues such as under-flagged on veiled forms of hate that perpetuate stigma and over-flagging of slurs that have been reclaimed by marginalized communities.

An Interactive Exploratory Tool for the Task of Hate Speech Detection

A suite of interactive modules to support the exploration of various aspects of this technology, and particularly of those components that rely on English models and datasets for hate speech detection, a subtask within ACM are described.

Fluent: An AI Augmented Writing Tool for People who Stutter

Fluent is presented, an AI augmented writing tool which assists people who stutter in writing scripts which they can speak more fluently and can be beneficial for certain important life situations like giving a talk, presentation, etc.



BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

A new language representation model, BERT, designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers, which can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks.

Attention is All you Need

A new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely is proposed, which generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data.

Designing User Interface Elements to Improve the Quality and Civility of Discourse in Online Commenting Behaviors

Exposure to CAPTCHAs featuring image sets previously validated to evoke low-arousal positive emotions significantly increased the positivity of sentiment and the levels of complexity and social connectedness in participants' posts.

The treatment of ties in ranking problems.

#thyghgapp: Instagram Content Moderation and Lexical Variation in Pro-Eating Disorder Communities

It is found that the pro-ED community has adopted non-standard lexical variations of moderated tags to circumvent these restrictions and express more toxic, self-harm, and vulnerable content.

Reconsidering Community Self-Moderation: the Role of Research in Supporting Community-Based Models for Online Content Moderation

Research in online content moderation has a long history of exploring different forms that moderation can take, including both user-driven moderation models on community-based platforms like

Interventions to counter hate speech

There is limited evidence on the effectiveness of interventions to counter hate speech. There is a lack of rigorous impact evaluations in this area and those that do exist tend to focus on individual

The philosophical basis of algorithmic recourse

It is argued that two essential components of a good life - temporally extended agency and trust - are underwritten by recourse, and a revised approach to recourse is suggested.

"Why is 'Chicago' deceptive?" Towards Building Model-Driven Tutorials for Humans

It is found that tutorials indeed improve human performance, with and without real-time assistance, and although deep learning provides superior predictive performance than simple models, tutorials and explanations from simple models are more useful to humans.