LAUSR

Common Sense knowledge bases and models have been shown to embed bias. We investigate the source of such bias in a knowledge model called common sense transformer (COMET) by training it on various combinations of language models and knowledge bases. We experiment with three language models of different sizes and architectures, and two knowledge bases with different modeling principles. We use sentiment and regard as proxy measures of bias and analyze bias using three methods: overgeneralization and disparity, keyword outliers, and relational dimensions. Our results show that larger models tend to be more nuanced in their biases but are more biased than smaller models in certain categories (e.g., utility of religions), which can be attributed to the larger knowledge accumulated during pretraining. We also observe that training on a larger set of common sense knowledge typically leads to more bias, and that models generally have stronger negative regard than positive.