A Systematic Study of Gender and Religion Bias in Stories

Bias is and will always be a topic of interest in Natural Language Applications. The stereotypes that are learnt by the language models do have a severe effect when used in the real world. Specifically in creative applications such as story generation where the user directly interacts with the model and its generated text. With this project, we investigate the extent of gender and religion bias learnt by word embedding models, whether there is a change in bias over a period of time and finally what effect does finetuning large pre-trained models affect bias. Project Gutenberg and OCLC datasets are used to experiment. Methods such as ranked lists and WEAT are used to detect bias in the word2vec model. The generated text is evaluated by automatic and manually to understand the effect of bias. The results suggest that bias does not change significantly over time and finetuning does reduce bias in the generated text.