Towards Fair and Selectively Privacy-Preserving Models Using Negative Multi-Task Learning (Student Abstract)

Liyuan Gao, Huixin Zhan, Austin Chen, Victor S. ShengAAAI 2023 · DOI

TL;DR

This paper introduces a novel "negative multi-task learning" framework to simultaneously improve model utility and reduce gender bias in NLP tasks. By training a model to perform a primary task while actively trying to fail at predicting gender, they achieve fairer models without sacrificing performance.

What Problem Does This Paper Address?

Deep learning models excel at processing vast amounts of text data, but this capability comes with two significant challenges: privacy leakage and social bias. When trained on real-world text, models can inadvertently learn and even amplify existing biases like sexism and racism. Furthermore, sensitive personal information embedded within the training data can be leaked through the model's predictions. This paper tackles both issues by proposing a method to protect sensitive information and simultaneously mitigate gender bias in NLP models.

Methods

The authors propose a two-pronged approach:

  1. Selective Privacy-Preserving Method: This technique focuses on obscuring only individuals' sensitive information within word embeddings. They define a function to detect sensitive words (like names, gendered terms, etc.) and then apply a perturbation function to their embeddings, adding normally distributed noise. This aims to protect privacy without drastically altering the word's meaning for the main task.
  2. Negative Multi-Task Learning Framework: This is the core innovation. It involves training a model with two objectives:
  3. Main Task Classification: The primary goal, e.g., sentiment analysis or medical text classification.
  4. Gender Prediction Task (with Negative Loss): The model is trained to predict gender, but the loss for this task is subtracted from the total loss (hence "negative"). The idea is to penalize the model for learning gender-specific features, forcing it to become worse at predicting gender.

    The final loss is formulated as Loss = L_main-task - λ * L_gender-prediction, where λ is a hyperparameter controlling the trade-off between main task accuracy and gender bias reduction. The goal is to improve the main task's accuracy while simultaneously decreasing the accuracy of the gender prediction task.

Datasets

The paper evaluates their approach on two types of tasks:

  • Sentiment Analysis: While not explicitly naming a dataset, they mention evaluating on "negative sentiment" and "positive sentiment" across four groups, implying variations or subsets of sentiment analysis datasets.
  • Medical Text Classification: Similar to sentiment analysis, a specific dataset name is not provided, but the task is clearly defined.

Key Insights and Ideas

  • "Unlearning" Bias: The negative multi-task learning approach is a clever way to force models to "unlearn" or avoid learning gender-specific features by actively penalizing them for learning them.
  • Trade-off Management: The λ parameter in the loss function provides a tunable knob to balance the desire for high accuracy on the main task with the need for fairness.
  • Gender Bias in Sentiment: The study found that sentiment analysis models were better at identifying sentiment from females than males, and that positive sentiment exhibited higher gender bias than negative sentiment. The authors hypothesize this is because females might use more distinct positive words, making them easier to classify.

Key Results

  • Medical Text Classification: The negative multi-task learning framework with selective privacy-preserving achieved the lowest disparity score (indicating reduced gender bias) and the highest accuracy on the medical text classification task.
  • Sentiment Analysis: The paper presents disparity scores for sentiment analysis, showing that their negative multi-task learning approach can reduce gender bias compared to single-task learning models. GloVe embeddings generally performed better than Word2vec.

Limitations and Future Work

  • Focus on Gender: The current work specifically addresses gender bias. The authors acknowledge that other biases (age, race, etc.) also exist and plan to extend their negative multi-task learning framework to mitigate multiple biases simultaneously.
  • Dataset Specificity: While the results are promising, the lack of specific dataset names makes it harder to replicate or compare directly with other works on those exact benchmarks.
  • Privacy Method Details: The privacy-preserving method is described as adding noise, but the effectiveness and potential impact on utility beyond what's shown in the results are not deeply explored.

Summary updated

BibTeX
@inproceedings{gao2023towards,
  author = {Liyuan Gao and Huixin Zhan and Austin Chen and Victor S. Sheng},
  title = {{Towards Fair and Selectively Privacy-Preserving Models Using Negative Multi-Task Learning (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i13.26967},
}
Figure