Bigger Data or Fairer Data? Augmenting BERT via Active Sampling for Educational Text Classification

Proceedings of the 29th International Conference on Computational Linguistics, pages 1275–1285 October 12–17, 2022. 1275 Bigger Data or Fairer Data? Augmenting BERT via Active Sampling for Educational Text Classification Lele Sha, Yuheng Li, Dragan Gaševi´c and Guanliang Chen∗ Centre for Learning Analytics, Monash University Melbourne, Victoria, Australia {lele.sha1, yuheng.li, Dragan.Gasevic, Guanliang.Chen}@monash.edu Abstract Pretrained Language Models (PLMs), though popular, have been diagnosed to encode bias against protected groups in the representations they learn, which may harm the prediction fair- ness of downstream models. Given that such bias is believed to be related to the amount of demographic information carried in the learned representations, this study aimed to quantify the awareness that a PLM (i.e., BERT) has re- garding people’s protected attributes and aug- ment BERT to improve prediction fairness of downstream models by inhibiting this aware- ness. Specifically, we developed a method to dynamically sample data to continue the pre- training of BERT and enable it to generate rep- resentations carrying minimal demographic in- formation, which can be directly used as in- put to downstream models for fairer predic- tions. By experimenting on the task of clas- sifying educational forum posts and measur- ing fairness between students of different gen- der or first-language backgrounds, we showed that, compared to a baseline without any ad- ditional pretraining, our method improved not only fairness (with a maximum improvement of 52.33%) but also accuracy (with a maximum improvement of 2.53%). Our method can be generalized to any PLM and demographic at- tributes. All the codes used in this study can be accessed via https://github.com/ lsha49/FairBERT_deploy. 1 Introduction Pretrained Language Models (PLMs) have been increasingly applied to tackle various NLP tasks in recent years (Li et al., 2019; Yoon et al., 2019; Chan and Fan, 2019; Araci, 2019; Zhang et al., 2019). Along with the wide application of PLMs is grow- ing concerns about the bias encoded in the represen- tations generated by these PLMs (Jin et al., 2020; Lu et al., 2020). For instance, de Vassimon Manela ∗ Corresponding author. et al. (2021) demonstrated that stereotypical asso- ciations were encoded in PLMs when tackling a pronoun resolution task. More importantly, such bias has been demonstrated to be harmful to the pre- diction fairness of downstream models, i.e., there exists a consistent gap between the prediction ac- curacy for people of different protected attributes. For example, Minot et al. (2021) showed that the gender-related bias embedded in PLMs could be propagated to downstream classification tasks in medical scenarios. As a remedy, researchers have endeavored to develop techniques to debias PLMs. These tech- niques, more often than not, focused on the fine- tuning stage when using a PLM, e.g., correcting the bias hidden behind the learned representations by removing associations between embedding features and protected attributes during the fine-tuning pro- cess or using a protected-attribute-balanced dataset to fine-tune a PLM (de Vassimon Manela et al., 2021). It is worth noting that, though assuming that the bias contained in a PLM is associated with the amount of …

Content truncated to 500 words