Olatz Perez de Viñaspre: The effect of social biases in language models

published Oct 05, 2023

Keynote talk by Olatz Perez de Viñaspre at Plone conference 2023, Eibar, Basque Country.

Current language models are trained on huge amounts of texts. The quality and content of such text has a direct effect on the new generations created by the language model. In this talk we will focus on how language models reproduce the biases present in society.

Models are big. For the Falcon model you need 400 GB on your computer.
Masked language models predict hidden words in sentences. ChatGPT is a Causal Language Model, or generative model. That is currently the biggest part of the language models evolutionary tree.
The language corpus of ChatGPT is roughly 90 percent English. German is 0.17 percent. So what happens, is that the models have an Anglo-centric bias.
Most systems are proprietary, not open source. Meta had the Llama model, with 68% performance compared to ChatGPT. It went open source and three weeks later it was more than 90%.
Bias: allocational or representational harm. In one model, black persons were sometimes recognised as monkeys because the model had not been trained well enough on faces of black persons. Also, calling a woman "independent" is positive, but it is still a bias: you don't often call a man independent.
How do you measure bias? Manually made datasets often contain problems: they are biased themselves. Recently a more objective solution: Marked Personas. Ask a question like "define a white male" and compare the answer with "define a black woman". Does this show biases?
Where do models get their data? Web, books, videos. But the internet is also full of hate speech, so you can train a model on hate. So there is a problem of quantity versus quality. There is a lot of effort on debiasing models, but is still an open task with many edges. Ethics will need to guide the further development of large language models.
Different languages can have different biases. For example Basque has no he or she. And we see that if you translate "he/she is a doctor" to English, it becomes "he", and for a nurse it becomes "she". So you see bias in translation.