Debiasing a Corpus with Semi-Supervised Topic Modelling : Swapping Gendered Words to Reduce Potentially Harmful Associations

Gender biases are present in many NLP models. Such biased models can have large negative consequences on individuals. This work is a case study where we attempt to reduce them in a corpus consisting of Wikipedia articles about persons, in order to reduce them in models that would be trained on the corpus. For this, we apply two methods of modifying the corpus’s documents. Both methods replace gendered words (such as ‘mother’, ‘father’ and ‘parent’) with each other to change the contexts in which they each appear. The analysis and comparison of those two methods show that one of them is indeed suited to reduce gender biases. By modifying 35% of the corpus’s documents, the context of gendered words seems equal between the three considered genders (feminine, masculine and non-binary). This is confirmed through the performance of coreference resolution models trained with word embeddings fine-tuned on the corpus before and after modifying it. Evaluating these models on schemas specifically designed to point out gender biases in coreference resolution models shows that the model using the modified corpus is indeed less gender biased than the original. Our analysis further shows that the method does not compromise the corpus’s overall quality.