"Tackling bias in text classification with responsible AI"

• Summary: Bias in language has attracted a lot of recent interest with many studies exposing a large number of offensive associations related to gender and race on publicly available word embeddings (Bolukbasi et al, 2016) as well as how these associations have evolved over time (Kutuzov et al, 2018). In this thesis, we consider to use explanation methods to investigate any inappropriate attention that might has been paid to protected group identifiers (related to e.g., gender or race) and might lead to bias on certain groups. Furthermore, the explanation method will contribute to the training phrase of the model in order to avoid placing (wrong) attention to protected group identifiers.
• Related work:
  • Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings
  • Diachronic word embeddings and semantic shifts: a survey
  • Contextualizing Hate Speech Classifiers with Post-hoc Explanation,
  • Measuring and Mitigating Unintended Bias in Text Classification

Reducing Gender Bias in Abusive Language Detection

Advisors: Prof. Dr. Eirini Ntoutsi, Yi Cai