Reinforcement Guided Multi-Task Learning Framework for Low-Resource Stereotype Detection

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 6703 - 6712 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Reinforcement Guided Multi-Task Learning Framework for Low-Resource Stereotype Detection Rajkumar Pujari1 and Erik Oveson2 and Priyanka Kulkari2 and Elnaz Nouri3 1Purdue University ∗ 2Microsoft, Redmond 3Microsoft Research, Redmond rpujari@purdue.edu {erikov,priyak,elnouri}@microsoft.com Abstract As large Pre-trained Language Models (PLMs) trained on large amounts of data in an unsuper- vised manner become more ubiquitous, iden- tifying various types of bias in the text has come into sharp focus. Existing ‘Stereotype Detection’ datasets mainly adopt a diagnostic approach toward large PLMs. Blodgett et al. (2021a) show that there are significant reliabil- ity issues with the existing benchmark datasets. Annotating a reliable dataset requires a precise understanding of the subtle nuances of how stereotypes manifest in text. In this paper, we annotate a focused evaluation set for ‘Stereo- type Detection’ that addresses those pitfalls by de-constructing various ways in which stereo- types manifest in text. Further, we present a multi-task model that leverages the abundance of data-rich neighboring tasks such as hate speech detection, offensive language detection, misogyny detection, etc., to improve the empiri- cal performance on ‘Stereotype Detection’. We then propose a reinforcement-learning agent that guides the multi-task learning model by learning to identify the training examples from the neighboring tasks that help the target task the most. We show that the proposed models achieve significant empirical gains over exist- ing baselines on all the tasks. 1 Introduction Following Peters et al. (2018) and then Vaswani et al. (2017), several empirically effective large pre-trained language models (PLMs) (Devlin et al., 2019; Yang et al., 2019; Liu et al., 2019; Lewis et al., 2020; Brown et al., 2020) have been pro- posed. Their empirical success on several existing tasks has resulted in them being ubiquitously used in many NLP applications that interact with hu- mans on a daily basis (Olteanu et al., 2020). As they are trained in an unsupervised manner on large amounts of arbitrary web data, harmful language ∗This work is a part of summer internship at Microsoft Research, Redmond and bias creeps into their output. This, in turn, is reflected in the applications that are built on top of them and further propagated in society through the use of those applications (Berk, 2017). This issue motivates the need for a two-pronged solution: 1) to diagnose and de-noise the bias in the PLMs and 2) to identify and regulate harmful text externally at the output level. Progress on the first front is necessary to de-bias the PLMs and prevent subtle biases that creep in through their distributed repre- sentations which can often only be recognized in retrospect. This results in deep distrust of such sys- tems among general population. The second front is paramount to detect and potentially block harm- ful output and also ensure that future iterations and novel proposals of the language models don’t take spurious paths to gaming the de-biasing techniques. In …

Content truncated to 500 words