The use of machine learning algorithms for the detection of abusive language online has a massive impact on the timely and proactive removal of harmful content from the Web. However, such systems are far from perfect. One critical bottleneck to-wards the development of more effective technologies is the lack of proper evaluation procedures. If unreliable systems achieve astonishing scores with traditional metrics, how do we recognise progress when we see it? The workshop on Novel Evaluation Approaches for Text Classification Systems aims to encourage a wider conversation about how misinformation detection, abusive language detection, sentiment analysis and related tasks should be evaluated. It promotes the design of new evaluation metrics and benchmarks that better capture a system's performance, and the development of new systems that perform better on these novel metrics. The workshop was held for the second time this year. This introduction to the workshop proceedings provides a summary of contributions.

PDF