Large pre-trained language models contain human-like biases of what is right and wrong to do

Do Multilingual Language Models Capture Differing Moral Norms?

“…Summarised, these observations already confirm the results of [14] in a larger multilingual setting and indicate that multilingual LMs indeed capture moral norms. To what extent they differ, however, is still unclear.…”

“…From initial experiments, a version of XLM-R tuned with the S-BERT framework [13] 1 shows good correlation with the global user study conducted by [14] when used with their MoralDirection (MD) framework (see Table 1). Simply mean-pooling representations from XLM-R [4], mBERT [6] 1.…”

“…and social norms [9,14]. Hereby the knowledge of a PLM is acquired during the self-supervised pre-training phase, which in case of most current state-of-the-art models uses scraped data from the web.…”

“…Therefore, approaches investigating the agreement of the model's norms and human values [14] and benchmark datasets [7] aiming to align human and machine values with human labelled data arose. The work of Jiang et al [10] showed promise in terms of obtaining such alignment, i.e., teaching these kinds of models commonsense moral reasoning.…”

“…Recall that the model is trained so that two sentences in different languages with the same semantics receive very similar embeddings. Also note that the tested statements provided by [14] do not aim to grasp cultural differences. An interesting question raised by these initial results is whether language alignment is in fact desirable when considering moral norms, which can differ in differing cultures.…”