ChatGPT Based Data Augmentation for Improved Parameter-Efficient Debiasing of LLMs

Dear Reviewer,

Thank you for your detailed assessment and recognition of our efforts. We would like to address the specific points mentioned:

1. Potential Introduction of Biases: We acknowledge that even with anti-stereotype prompts, ChatGPT can introduce nuanced biases or implicit toxicity, which may remain undetected despite our careful inspection (see Figure 4, Table 1, and Appendices). However, we highlight the following:

   • Effectiveness: The same synthetic dataset (synthetic-general) is effective at debiasing across several bias categories in different datasets, suggesting it doesn’t introduce new biases while mitigating others.
   • Comparison with Real-World Data: Real-world data collection is challenging and largely dependent on human input, which can vary in reliability and bias. In contrast, our synthetic data, generated with a thoughtfully designed prompting strategy, is more efficient, consistent, and offers a higher degree of control.
   • Area for Improvement: A potential additional check that might be added is another expert LLM to check for toxicity, using models like ToxicBERT or Self-Detoxification.

2. Experimental Setup Clarifications:

   • Limited Resources and Adapter Layers: Our experiments on Google Colab use default settings with an adapter layer after each transformer layer (12 layers for bert-base-uncased). We will include these details in the paper.
   • Full Finetuning: Full-finetuning/retraining is computationally expensive and may lead to overfitting given our dataset is small. Additionally, the work by Xie and Lukasiewicz, which we compare with, shows that parameter-efficient methods are better. We will discuss this in the paper.

3. Distribution Differences and Tailored Data Generation:

   • Distribution Differences: Since distributions differ between smaller and larger models, as well as between GPT and non-GPT families, it’s essential to tailor data generation. Our loss-guided prompting strategy (described in Section 3) addresses this. We generate data normally, calculate the loss using model M (the model being debiased), and use lower-loss data as few-shot examples to generate more in-distribution data, reducing the distribution gap.
   • Alternative Methods: We will discuss self-instruct and self-aligned methods as other possible alternatives.

We hope our explanations address your concerns. Please let us know if there are any further questions or suggestions. Thank you!

Dear Reviewer,

We sincerely appreciate your thoughtful review and the recognition of our paper’s contributions. We would like to respond to the concerns raised.

• Regarding the first point, while synthetic data generation and augmentation are not novel concepts, our approach to utilizing them for effective debiasing is innovative in this context. The method introduced in the work by Yu et al. [1] focuses on generating more diverse, high-quality, and unbiased data by controlling various attributes such as length, style, and location. Their approach prompts the model to generate data considering multiple dimensions (e.g., [Generate A, consider Dimension1, Dimension2, Dimension3, etc.]). In contrast, our method specifically leverages the model’s knowledge of social bias and social groups to generate anti-stereotyped data, aiming to counteract bias. Our attributes refer to names and entities within social groups, and our [Sentence, Target, Attribute] format is designed to support subsequent pipelines for effective debiasing via parameter-efficient fine-tuning. In summary, while Yu et al. [1]’s method uses prompting strategies to create diverse and unbiased synthetic datasets, our method actively leverages ChatGPT’s knowledge of social bias to generate targeted anti-stereotyped data for debiasing. We differ in goals and approaches.

• Regarding the second point, we apologize for any confusion. Current popular debiasing methods for LLMs generally involve 1) fine-tuning or retraining LLMs on specific datasets. 2) applying post-hoc mathematical operations on the frozen representations of a LM. The first two paragraphs in Section 2.2 of our paper introduce various attempts to build high-quality datasets for fine-tuning or retraining models. In the last paragraph, we mentioned the work by Xie and Lukasiewicz, which provides a comprehensive evaluation of different major parameter-efficient fine-tuning and model retraining methods for debiasing purposes. Their work also includes post-hoc methos like SentenceDebias (Liang et al., 2020) and SelfDebias (Schick et al., 2021). This work serves as our direct comparison and baselines.

Thank you once again for recognizing our potential and providing valuable suggestions. We hope these clarifications address your concerns, and we are happy to answer any further questions and incorporate these clarificaitons into the paper.

Dear Reviewer,

Thank you for your recognition and constructive comments. We appreciate your feedback and would like to address the concerns raised.

Regarding the technical novelty, we would like to emphasize that our contribution extends beyond simply the use of effective prompting. We present a holistic and easy-to-use pipeline that integrates knowledge about social bias specification, synthetic data generation in a specific format of [Sentence, Target, Attribute] that supports debiasing, and the associated adapter-based lightweight fine-tuning.

We propose distinct strategies for synthetic data generation, each with unique strengths. Our experiments with loss-guided prompting further improve our approach’s effectiveness and applicability. This process requires an understanding of bias specification, effective adaptation of LLM capabilities, and a thoughtfully designed pipeline that combines these elements. These efforts enable both domain and non-domain experts to quickly and efficiently generate synthetic data to debias models with minimal cost.

For more in-depth analysis and case studies, we included various levels of analysis and cases in the paper. Key experiments already included:

• Reducing Bias: Effectiveness in reducing bias in BERT and four GPT models across two datasets (Figures 1 and 2) with additional analysis on a third dataset capturing different biases (Table 5).
• Impact on Language Capabilities: Analyzing debiasing impact on model language capabilities (Tables 2-4).
• Data Similarity Analysis: Comparing our synthetic datasets to existing datasets to address memorization and data leakage (Section 6 and Table 6).
• Example Cases: Analysis of generated terms and sentences in our synthetic data (Figure 4, Table 1, and 7).
• Loss-Guided Technique: Testing the hypothesis on the importance of in-domain synthetic data generation (Section 3 “Loss-Guided Prompting”).
• Trade-Off Analysis: Evaluating the trade-off between language capabilities and debiasing performance, and performance with different training data sizes (Figures 5 and 6).

We hope this extensive set of experiments provides sufficient evidence of our approach’s effectiveness. We would also appreciate any concrete suggestions for additional experiments to strengthen our results.

Thank you once again for your constructive feedback! We found it to be very helpful. Feel free to let us know if there are any further concerns or if any additional adjustments are needed.