Sustainable Modular Debiasing of Language Models

Introduction

Recent work has shown that pretrained language models such as ELMo (Peters et al., 2018), BERT (Devlin et al., 2019), or GPT-2 (Radford et al., 2019) tend to exhibit a range of stereotypical societal biases, such as racism and sexism (e.g., Kurita et al., 2019; Dev et al., 2020; Webster et al., 2020; Nangia et al., 2020; Barikeri et al., 2021, inter alia). The reason for this lies in the distributional nature of these models: human-produced corpora on which these models are trained are abundant with stereotypically biased concept co-occurrences (for instance, male terms like man or son appear more often together with certain career terms like doctor or programmer than female terms like women or daughter) and the PLMs models, being trained with language modeling objectives, consequently encode these biased associations in their parameters. While this effect can lend itself to diachronic analysis of societal biases (e.g., Garg et al., 2018; Walter et al., 2021), it represents stereotyping, one of the main types of representational harm (Blodgett et al., 2020) and, if unmitigated, may cause severe ethical issues in various sociotechnical deployment scenarios.

To alleviate this problem and ensure fair language technology, previous work introduced a wide range of bias mitigation methods (e.g., Bordia and Bowman, 2019; Dev et al., 2020; Lauscher et al., 2020a, inter alia). All existing debiasing approaches, however, modify all parameters of the PLMs which has two prominent shortcomings: (1) it comes with a high computational costWhile a full fine-tuning approach to PLM debiasing may still be feasible for moderate-sized PLMs like BERT Devlin et al. (2019), it is prohibitively computationally expensive for giant language models like GPT-3 Brown et al. (2020) or GShard Lepikhin et al. (2020). and (2) can lead to (catastrophic) forgetting (McCloskey and Cohen, 1989; Kirkpatrick et al., 2017) of the useful distributional knowledge obtained during pretraining. For example, Webster et al. (2020) incorporate counterfactual debiasing already into BERT’s pretraining: this implies a debiasing framework in which a separate “debiased BERT” instance needs to be trained from scratch for each individual bias type and specification. In sum, current debiasing procedures designed for pretraining or full fine-tuning of PLMs have a large carbon footprint Strubell et al. (2019) and consequently jeopardize the sustainability Moosavi et al. (2020) of fair representation learning in NLP.

In this work, we move towards more sustainable removal of stereotypical societal biases from pretrained language models. To this end, we propose Adele (Adapter-based DEbiasing of LanguagE Models), a debiasing approach based on the the recently proposed modular adapter framework (Houlsby et al., 2019; Pfeiffer et al., 2020a). In Adele, we inject additional parameters, the so-called adapter layers into the layers of the PLM and incorporate the “debiasing” knowledge only in those parameters, without changing the pretrained knowledge in the PLM. We show that, while being substantially more efficient (i.e., sustainable) than existing state-of-the-art debiasing approaches, Adele is just as effective in bias attenuation.

The contributions of this work are three-fold: (i) we first present Adele, our novel adapter-based framework for parameter-efficient and knowledge-preserving debiasing of PLMs. We combine Adele with one of the most effective debiasing strategies, Counterfactual Data Augmentation (CDA; Zhao et al., 2018), and demonstrate its effectiveness in gender-debiasing of BERT Devlin et al. (2019), the most widely used PLM. (ii) We benchmark Adele in what is arguably the most comprehensive set of bias measures and data sets for both intrinsic and extrinsic evaluation of biases in representation spaces spanned by PLMs. Additionally, we study a previously neglected effect of fairness forgetting present when debiased PLMs are subjected to large-scale downstream training for specific tasks (e.g., natural language inference, NLI); we show that Adele’s modular nature allows to counter this undesirable effect by stacking a dedicated task adapter on top of the debiasing adapter. (iii) Finally, we successfully transfer Adele’s debiasing effects to six other languages in a zero-shot manner, i.e., without relying on any debiasing data in the target languages. We achieve this by training the debiasing adapter stacked on top of the multilingual BERT on the English counterfactually augmented dataset.

Adele: Adapter-Based Debiasing

In this work, we seek to fulfill the following three desiderata: (1) we want to achieve effective debiasing, comparable to that of existing state-of-the-art debiasing methods while (2) keeping the training costs of debiasing significantly lower; and (3) fully preserving the distributional knowledge acquired in the pretraining. To meet all three criteria, we propose debiasing based on the popular adapter modules (Houlsby et al., 2019; Pfeiffer et al., 2020a). Adapters are lightweight neural components designed for parameter-efficient fine-tuning of PLMs, injected into the PLM layers. In downstream fine-tuning, all original PLM parameters are kept frozen and only the adapters are trained. Because adapters have fewer parameters than the original PLM, adapter-based fine-tuning is more computationally efficient. And since fine-tuning does not update the PLM’s original parameters, all distributional knowledge is preserved.

The debiasing adapters could, in principle, be trained using any of the debiasing strategies and training objectives from the literature, e.g., via additional debiasing loss objectives Qian et al. (2019, inter alia); Bordia and Bowman (2019, inter alia); Lauscher et al. (2020a, inter alia) or data-driven approaches such as Counterfactual Data Augmentation (Zhao et al., 2018). For simplicity, we opt for the data-driven CDA approach: it has been shown to offer reliable debiasing performance Zhao et al. (2018); Webster et al. (2020) and, unlike other approaches, it does not require any modifications of the model architecture nor training procedure.

In this work, we employ the simple adapter architecture proposed by Pfeiffer et al. (2021), in which only one adapter module is added to each layer of the pretrained Transformer, after the feed-forward sub-layer. The more widely used architecture of Houlsby et al. (2019) inserts two adapter modules per Transformer layer, with the other adapter injected after the multi-head attention sublayer. We opt for the “Pfeiffer architecture” because in comparison with the “Houlsby architecture” it is more parameter-efficient and has been shown to yield slightly better performance on a wide range of downstream NLP tasks Pfeiffer et al. (2020a, 2021). The output of the adapter, a two-layer feed-forward network, is computed as follows:

In our case, we train the adapters for debiasing: we inject adapter layers into BERT Devlin et al. (2019), freeze the original BERT’s parameters, and run a standard debiasing training procedure – language modeling on counterfactual data (§2.2) – during which we only tune the parameters of the debiasing adapters. At the end of the debiasing training, the debiasing functionality is isolated into the adapter parameters. This not only preserves the distributional knowledge in the Transformer’s original parameters, but also allows for more flexibility and “on-demand” usage of the debiasing functionality in downstream applications. For example, one could train a separate set of debiasing adapters for each bias dimension of interest (e.g., gender, race, religion, sexual orientation) and selectively combine them in downstream tasks, depending on the constraints and requirements of the concrete sociotechnical environment.

2 Counterfactual Augmentation Training

In the context of representation debiasing, counterfactual data augmentation (CDA) refers to the automatic creation of text instances that in some way counter the stereotypical bias present in the representation space. CDA has been successfully used for attenuating a variety of bias types, e.g., gender and race, and in several variants, e.g., with general terms describing dominant and minoritized groups, or with personal names acting as proxies for such groups Zhao et al. (2018); Lu et al. (2020). Most commonly, CDA modifies the training data by replacing terms describing one of the target groups (dominant or minoritized) with terms describing the other group. Let $S$ be our training corpus, consisting of sentences $s$ and let $T=\{(t_{1},t_{2})^{i}\}_{i=1}^{N}$ be a set of $N$ term pairings between the dominant and minoritized group (i.e., $t_{1}$ is a term representing the dominant group, e.g., man, and $t_{2}$ is a corresponding term representing the minoritized group, e.g., woman). For each sentence $s_{i}$ and each pair $(t_{1},t_{2})$, we check whether either $t_{1}$ or $t_{2}$ occur in $s$: if $t_{1}$ is present, we replace its occurrence with $t_{2}$ and vice versa. We denote the counterfactual sentence of $s$ obtained this way with $s^{\prime}$ and the whole counterfactual corpus with $S^{\prime}$. We adopt the so-called two-sided CDA from (Webster et al., 2020): the final corpus for debiasing training consists of both the original and counterfactually created sentences. Finally, we train the debiasing adapter via masked language modeling on the counterfactually augmented corpus $S\cup S^{\prime}$. We train sequentially by first exposing the adapter to the original corpus $S$ and then to the augmented portion $S^{\prime}$.

Experiments

We showcase Adele for arguably the most explored societal bias – gender bias – and the most widely used PLM, BERT. We profile its debiasing effects with a comprehensive set of intrinsic and downstream (i.e., extrinsic) evaluations.

We test Adele on three intrinsic (BEC-Pro, DisCo, WEAT) and two downstream debiasing benchmarks (Bias-STS-B and Bias-NLI). We now describe each of the benchmarks in more detail.

We intrinsically evaluate Adele on the BEC-Pro data set (Bartl et al., 2020), designed to capture gender bias w.r.t. professions. The data set consists of 2,700 sentence pairs in the format (“m [temp] p”; “f [temp] p”), where m is a male term (e.g., boy, groom), f is a female term (e.g., girl, bride), p is a profession term (e.g., mechanic, doctor), and [temp] is one of the predefined connecting templates, e.g., “is a” or “works as a”.

We measure the bias on BEC-Pro using the bias measure of Kurita et al. (2019). They compute the association $a_{t,p}$ between a gender term $t$ (male or female) and a profession $p$ as:

where $P(t)_{t}$ is the probability of the PLM generating the target term $t$ when only $t$ itself is masked, and $P(t)_{t,p}$ is the probability of $t$ being generated when both $t$ and the profession $p$ are masked. The bias score $b$ is then simply a difference in the association score between the male term $m$ and its corresponding female term $f$: $b=a_{m,p}-a_{f,p}$. We measure the overall bias on the whole dataset in two complementary ways: (a) by averaging the bias scores $b$ across all 2,700 instances ($\varnothing$ bias) and (b) by measuring the percentage of instances for which $b$ is below some threshold value: we report this score for two different thresholds ($0.1$ and $0.7$).

Bartl et al. (2020) additionally published a German version of the BEC-Pro data set, which we use to evaluate Adele’s zero-shot transfer abilities.

Discovery of Correlations (DisCo).

The second data set for intrinsic debiasing evaluation, DisCo (Webster et al., 2020), also relies on templates (e.g., “[PERSON] studied [BLANK] at college”). For each template, the [PERSON] slot is filled first with a male and then with a female term (e.g., for the pair (John, Veronica), we get John studied [BLANK] at college and Veronica studied [BLANK] at college). Next, for each of the two instances, the model is asked to fill the [BLANK] slot: the goal is to determine the difference in the probability distribution for the masked token, depending on which term is inserted in the [PERSON] slot. While Webster et al. (2020) retrieve the top three most likely terms for the masked position, we retrieve all terms t with the probability $p(t)>0.1$.We argue that retrieving more terms from the distribution allows for a more accurate estimate of the bias.

Let $C^{(i)}_{m}$ and $C^{(i)}_{f}$ be the candidate sets obtained for the $i$-th instance when filled with a male [PERSON] term $m$ and the corresponding female term $f$, respectively. We then compute two different measures. The first is the average fraction of shared candidates between the two sets ($\varnothing\text{frac}$):

with $N$ as the total number of test instances. Intuitively, a higher average fraction of shared candidates indicates lower bias.

For the second measure, we retrieve the probabilities $p(t)$ for all candidates $t$ in the union of two sets $C^{(i)}=C^{(i)}_{m}\cup C^{(i)}_{f}$. We then compute the normalized average absolute probability difference:

We create test instances by collecting $100$ most frequent baby names for each gender from the US Social Security name statistics for 2019.https://www.ssa.gov/oact/babynames/limits.html We create pairs ($m$, $f$) from names at the same frequency rank in the two lists (e.g., Liam and Olivia). Finally, we remove pairs with ambiguous names that may also be used as general concepts (e.g., violet, a color), resulting in final $92$ pairs.

Word Embedding Association Test (WEAT).

As the final intrinsic measure, we use the well-known WEAT (Caliskan et al., 2017) test. Developed for detecting biases in static word embedding spaces, it computes the differential association between two target term sets $A$ (e.g., male terms) and $B$ (e.g., female terms) based on the mean (cosine) similarity of their embeddings with embeddings of terms from two attribute sets $X$ (e.g., science terms) and $Y$ (e.g., art terms):

The association $s$ of term $t\in A$ or $t\in B$ is computed as:

The significance of the statistic is computed with a permutation test in which $s(A,B,X,Y)$ is compared with the scores $s(A^{*},B^{*},X,Y)$ where $A^{*}$ and $B^{*}$ are equally sized partitions of $A\cup B$. We report the effect size, a normalized measure of separation between the association distributions:

where $\mu$ is the mean and $\sigma$ is the standard deviation.

Lauscher and Glavaš (2019) created XWEAT by translating some of the original WEAT bias specifications to six target languages: German (de), Spanish (es), Italian (it), Croatian (hr), Russian (ru), and Turkish (tr). We use their translations of the WEAT 7 gender test in the zero-shot debiasing transfer evaluation of adele.

Bias-STS-B.

The first extrinsic measure we use is Bias-STS-B, introduced by Webster et al. (2020), based on the well-known Semantic Textual Similarity-Benchmark (STS-B; Cer et al., 2017), a regression task where models need to predict semantic similarity for pairs of sentences. Webster et al. (2020) adapt STS-B for discovering gender-biased correlations. They start from neutral STS templates and fill them with a gendered term (man, woman) and a profession term from (Rudinger et al., 2018) (e.g., A man is walking vs. A nurse is walking and A woman is walking vs. A nurse is walking). The dataset consists of 16,980 such pairs. As a measure of bias, we compute the average absolute difference between the similarity scores of male and female sentence pairs, with a lower value corresponding to less bias. We couple the bias score with the actual STS task performance score (Pearson correlation with human similarity scores), measured on the STS-B development set.

Bias-NLI.

We select the task of understanding biased natural language inferences (NLI) as the second extrinsic evaluation. To this end, we fine-tune the original BERT as well as our adapter-debiased BERT on the MNLI data set (Williams et al., 2018). For evaluation, we follow Dev et al. (2020), and create a synthetic NLI data set that tests for the gender-occupation bias: it comprises NLI instances for which an unbiased model should not be able to infer anything, i.e., it should predict the neutral class. We use the code of Dev et al. (2020) and, starting from the generic template The a/an