BiasEdit: Debiasing Stereotyped Language Models via Model Editing

🔼 The figure illustrates the BiasEdit model, which uses lightweight editor networks to debias a language model. The editor networks generate parameter updates that specifically target stereotypical biases while preserving the model’s overall language modeling capabilities. This is achieved using two loss functions: a debiasing loss (Ld) which guides the networks to remove the bias by equalizing probabilities of stereotypical and anti-stereotypical contexts, and a retention loss (Lr) to ensure that unrelated associations are maintained. The figure shows that after editing, the model is debiased, and this debiasing effect is robust to gender reversal and semantic generality tests. The labels s, a, and m represent stereotyped, anti-stereotyped, and meaningless contexts respectively.

read the caption

Figure 2: Debiasing a language model with BiasEdit. Editor networks ϕitalic-ϕ\phiitalic_ϕ are trained to produce edit shifts on partial parameters 𝒲𝒲\mathcal{W}caligraphic_W of a language model while its parameters θ𝜃\thetaitalic_θ are frozen . After editing, an unbiased LM is obtained with the robustness of gender reversal and semantic generality. ℒdsubscriptℒ𝑑\mathcal{L}_{d}caligraphic_L start_POSTSUBSCRIPT italic_d end_POSTSUBSCRIPT and ℒrsubscriptℒ𝑟\mathcal{L}_{r}caligraphic_L start_POSTSUBSCRIPT italic_r end_POSTSUBSCRIPT refer to Equation 1 and 2 respectively. s: stereotyped. a: anti-stereotyped. m: meaningless.

🔼 This figure displays the Stereotype Score (SS) and the change in Language Modeling Score (ΔLMS) after applying bias editing to different layers of four distinct language models. The x-axis represents various combinations of blocks within the language model that were modified. ‘1’, ‘2’, and ‘3’ indicate edits to the first, second, and third blocks respectively. ‘12’ signifies edits to the first two blocks, and ‘123’ shows edits to the first three blocks. Negative numbers (-1, -2, -3) denote edits to the last, second-to-last, and third-to-last blocks respectively. ‘−21’ indicates edits to the last two blocks, and ‘−321’ represents edits to the last three blocks. The y-axis shows the SS and ΔLMS for each model, revealing the impact of the bias editing on both bias mitigation and overall language modeling capability.

read the caption

Figure 3: SS (%) and ΔΔ\Deltaroman_ΔLMS (%) of debiased language models after editing the last layer in the MLP of different blocks. 1/2/3: the first/second/third block. 12: the first 2 blocks. 123: the first 3 blocks. -1/-2/-3, the last/penultimate/antepenultimate block, -321: the last 3 blocks. -21: the last 2 blocks.

🔼 This figure demonstrates the robustness of the BiasEdit model to gender reversal. It displays the Stereotype Score (SS) for four different language models before debiasing (Pre-debias) and after debiasing with BiasEdit (Debiased). The gender reversal test set was created by reversing the gender terms in the original StereotypeSet dataset. A lower SS indicates less bias. The figure visually shows the impact of BiasEdit on reducing gender bias, even when presented with reversed gender terms, indicating that the model learns a more nuanced and equitable representation of gender.

read the caption

Figure 4: Gender Reversal Robustness. Pre-debias refers to SS of pre-trained language models on the gender reversal test set before debiasing. Debiased refers to SS of debiased models by BiasEdit.

🔼 Figure 5 presents a detailed analysis of gender bias in the GPT2-medium language model using bias tracing. Panel (a) compares the strength of bias associations across different layers of the model (hidden states, attention, and MLP layers). Panel (b) shows how bias associations vary depending on which specific words within the sentence are examined: the bias attribute word itself, the word preceding it, and the attribute term. Panels (c-h) further investigate the impact of these bias associations on the model’s output probabilities. This is done by selectively corrupting (adding noise to) the embeddings of bias-related words and then restoring different parts of the model’s internal activations (hidden states, attention, or MLP layers only). By comparing the changes in output probabilities before and after these manipulations, the figure illustrates how different parts of the model contribute to the overall gender bias.

read the caption

Figure 5: Gender bias tracing on GPT2-medium. (a) Comparing bias associations of bias attribute words on hidden states, attention layers, and MLP layers. (b) Comparing bias associations on single states of the bias attribute word, the token before the attribute term, and the attribute term. The bias impacts on output probability are mapped for the effect of (c-d) each hidden state on the context, (e-f) only MLP activations, and (g-h) only attention activations. * marks the corrupted bias attribute words and [] refers to the attribute terms in (c-h).

🔼 This figure presents the results of a bias tracing analysis performed on the GPT-2-medium language model, focusing on racial bias. It visualizes the association between bias and different components of the model’s architecture (hidden states, attention, and MLP layers) across various layers. The plots likely show how the model’s representation of race changes throughout the processing stages and how manipulating specific states within the model impacts the overall bias in its output. The subplots might show the effect of different interventions, such as removing or restoring specific activations.

read the caption

Figure 6: Race bias tracing on GPT2-medium.

🔼 This table presents the performance of different language models (LLMs) on a set of general benchmark tasks, both before and after bias mitigation. The metrics used are accuracies (in percentage). The models tested include Llama3, Mistral, Gemma, and GPT2-medium. The pre-edit and post-edit accuracies showcase the impact of the debiasing technique on the models’ general performance. The goal is to determine if bias editing affects the models’ ability to perform standard language understanding tasks.

read the caption

Table 2: Accuracies (%) of general model benchmarks. ’pre’: pre-edit, ‘post-’: post-edit, ‘GPT2m’: ‘GP2-medium’

🔼 This table compares the performance of BIASEDIT with and without the retention loss (ℒr). It shows the Stereotype Score (SS) and the change in Language Modeling Score (△LMS) for gender, race, and religion bias on GPT2-medium and Gemma-2b language models. The retention loss is designed to prevent negative impacts on the language modeling capabilities of the model during the debiasing process. By comparing the results with and without this loss, the table highlights its effectiveness in preserving the model’s overall performance while reducing bias.

read the caption

Table 3: BiasEdit w and w/o the retention loss ℒrsubscriptℒ𝑟\mathcal{L}_{r}caligraphic_L start_POSTSUBSCRIPT italic_r end_POSTSUBSCRIPT.

🔼 This table presents the Stereotype Score (SS) on a synonym-augmented test set. The synonym-augmented test set replaces attribute terms in the original StereoSet test set with their synonyms. This assesses the robustness and generalizability of the BIASEDIT debiasing method by evaluating its performance on semantically similar, but not identically worded, instances of bias.

read the caption

Table 4: SS (%) on the synonym-augmented test set.

🔼 This table presents the distribution of samples across different bias categories (gender, race, and religion) within the StereoSet dataset used in the study. It shows the number of samples used for training (Strain), development (Sdev), and testing (Stest) sets for each bias type.

read the caption

Table 5: The numbers of samples about different bias in our dataset.

🔼 This table presents the results of evaluating different bias mitigation methods on the Crowds-Pairs dataset. It shows the Stereotype Score (SS), a metric representing the percentage of times a model prefers stereotypical contexts over anti-stereotypical ones. Lower scores indicate less bias. The table compares the performance of four established baselines (CDA, SentenceDebias, Self-Debias, INLP) against the proposed method, BiasEdit, across three bias types (gender, race, religion) and four language models (GPT2-medium, Gemma-2b, Mistral-7B-v0.3, Llama3-8B). Pre-edit scores represent the bias levels before any mitigation.

read the caption

Table 6: Stereotype Score (%) for evaluating the baselines and BiasEdit on Crows-Pairs.

🔼 This table presents the results of training bias editor networks using two different approaches: one using data for a single bias type (gender, race, or religion) and another using a mixture of all three bias types. The table shows the stereotype score (SS) and the change in language modeling score (△LMS) for each model (GPT2-medium, Gemma-2b, Mistral-7B-v0.3, and Llama3-8B) and bias type, comparing the performance of the two training methods. Lower SS values indicate better debiasing performance, and △LMS close to 0 indicates minimal impact on the language modeling ability.

read the caption

Table 7: Training editor networks with data for one type of bias vs. mixed types of bias.