Never before was it so easy to write a powerful NLP system, never before did it have such a potential impact. However, these systems are now increasingly used in applications they were not intended for, by people who treat them as interchangeable black boxes. The results can be simple performance drops, but also systematic biases against various user groups.
In this talk, I will discuss several types of biases that affect NLP models (based on Shah et al. 2020 and Hovy & Spruit, 2016), what their sources are, and potential counter measures.
– bias stemming from data, i.e., selection bias (if our texts do not adequately reflect the population we want to study), label bias (if the labels we use are skewed), and semantic bias (the latent stereotypes encoded in embeddings).
– biases deriving from the models themselves, i.e., their tendency to amplify any imbalances that are present in the data.
– design bias, i.e., the biases arising from our (the practitioners) decisions which topics to explore, which data sets to use, and what to do with them.
As a consequence, we as NLP practitioners suddenly have a new role, in addition to researcher and developer: considering the ethical implications of our systems, and educating the public about the possibilities and limitations of our work. The time of academic innocence is over, and we need to address this newfound responsibility as a community.
For each bias, I will provide real examples and discuss the possible ramifications for a wide range of applications, and the various ways to address and counteract these biases, ranging from simple labeling considerations to new types of models.
– Deven Shah, H. Andrew Schwartz, & Dirk Hovy. 2020. Predictive Biases in Natural Language Processing Models: A Conceptual Framework and Overview. In Proceedings of ACL. [https://www.aclweb.org/anthology/2020.acl-main.468/]
– Dirk Hovy & Shannon L. Spruit. 2016. The Social Impact of Natural Language Processing. [https://www.aclweb.org/anthology/P16-2096.pdf]