Bias measurement and fairness techniques, privacy, environmental cost and open questions on the future of NLP
|
NLP - Ethics and Future Directions (pdf)
|
420.29 K |
Bias in language models has sources such as historical bias in the training data, and it is measured by template-based tests, the Word Embedding Association Test and counterfactual evaluation. In a template test, male and female completions of The [MASK] is a doctor are compared, and in counterfactual evaluation demographic attributes in a text are swapped and the change in predictions is measured. Fairness interventions are applied to the data (pre-processing), the training objective (in-processing) or the outputs (post-processing), followed by continuous monitoring. Privacy risks are addressed by differential privacy training and federated learning. The carbon cost of large models is weighed against accuracy gains, and future directions include multilingual models and open questions about alignment and artificial general intelligence.