Directions in abusive language training data, a systematic review: Garbage in, garbage out

Area of Science:

• Computational Linguistics
• Machine Learning
• Data Science

Background:

• Machine learning models for detecting abusive content (hate speech, harassment) are increasingly vital.
• The performance of these models heavily depends on the quality and characteristics of the training datasets.
• Developing comprehensive, unbiased, and theoretically-informed datasets is a significant challenge.

Purpose of the Study:

• To systematically review existing publicly available training datasets for abusive language classifiers.
• To report on the creation of a centralized resource for abusive language data.
• To discuss the implications of open science for dataset sharing in this domain.

Main Methods:

• Systematic literature review of 63 publicly available datasets.
• Creation and curation of a dedicated website (hatespeechdata.com) to catalogue datasets.
• Analysis of challenges, opportunities, and risks associated with open science and data sharing.

Main Results:

• Identification and cataloguing of 63 relevant training datasets.
• Discussion of the critical role of dataset quality in model performance (Garbage In, Garbage Out).
• Exploration of the dual nature of open science: benefits and socio-ethical risks.

Conclusions:

• High-quality, diverse, and unbiased datasets are crucial for effective abusive content detection.
• Open science and data sharing offer benefits but require careful ethical consideration.
• Evidence-based recommendations are provided for practitioners developing new abusive content training datasets.