Addressing Age-Related Bias in Sentiment Analysis

Mark Díaz, Isaac Johnson, Amanda Lazar, Anne Marie Piper, Darren GergleIJCAI 2019 · DOI

TL;DR

This paper reveals that popular sentiment analysis tools and word embeddings exhibit significant age-related bias, often scoring sentences with "young" adjectives more positively than those with "old" adjectives. The researchers demonstrate that this bias can be substantially reduced by carefully curating the training data, offering a practical approach to mitigate such issues in AI systems.

What Problem Does This Paper Address?

Sentiment analysis is a powerful tool used in various applications, from market analysis to political campaign strategies. However, like many AI systems, sentiment analysis tools can inadvertently perpetuate societal biases. This paper focuses on age-related bias, specifically how sentiment analysis algorithms and word embeddings might unfairly discriminate against certain age groups. The concern is that automated opinion polling could misrepresent attitudes towards age-related issues, leading to flawed decision-making. The authors aim to systematically analyze this bias, understand its origins, and propose a method to alleviate it.

Methods

The researchers employed a multi-phase approach:

  • Phase 1: Explicit Encoding of Age: They analyzed 15 sentiment analysis tools and 10 GloVe word embeddings by feeding them sentences where age-related terms ("old" vs. "young") were explicitly varied. This allowed them to directly measure how the tools responded to these explicit age cues. They used multinomial log-linear regression to analyze the sentiment outputs.
  • Phase 2: Implicit Encoding of Age: To investigate bias embedded within word embeddings, they generated "older" and "younger" analogs of common English adjectives using vector math on GloVe embeddings. These implicitly aged adjectives were then used in template sentences and analyzed by the sentiment tools. This phase aimed to uncover biases that are not explicitly stated but are learned from the underlying word representations.
  • Phase 3: Addressing Bias via Training Data: The team focused on a specific corpus-based sentiment analysis model (Sentiment140). They created custom models by manipulating the training data: one set included age-related tweets ("Age-Related"), another excluded them ("Age-Removed"), and a baseline used the original data ("Original"). This allowed them to pinpoint how bias in training data affects model output and to test a data-centric mitigation strategy.

Datasets

  • Blog Posts and Comments: The study sourced sentences from a community of older adult bloggers and their associated comments, providing a naturalistic context for age-related discussions.
  • Sentiment Analysis Tools: 15 popular sentiment analysis tools were evaluated.
  • GloVe Word Embeddings: 10 widely-used GloVe word embeddings were analyzed.
  • Sentiment140 Dataset: Used for Phase 3, this is a publicly available, annotated training dataset of 1 million tweets for corpus-based sentiment analysis. The researchers created custom subsets: "Age-Related" (13,781 tweets containing "young" or "old") and "Age-Removed" (the original dataset minus the "Age-Related" tweets).

Key Insights and Ideas

  • Age Bias is Pervasive: Both explicit age terms and implicitly aged word embeddings reveal a consistent bias in sentiment analysis tools, favoring "young" over "old."
  • Social Media Data Amplifies Bias: Sentiment analysis tools trained or validated on social media data (like Twitter) tend to exhibit more age-related bias compared to those trained on other sources like product reviews or news.
  • Word Embeddings Learn Stereotypes: The "implicit encoding" phase demonstrated that word embeddings can learn and propagate age-related stereotypes, even when the age term isn't explicitly stated but is derived through word vector manipulations.
  • Data Curation is a Powerful Mitigation Tool: The most significant finding is that simply removing age-specific examples from the training data can drastically reduce age-related bias in a sentiment classifier, suggesting that data preprocessing is a viable and effective strategy.

Key Results

  • Explicit Age Bias: Sentences with "young" adjectives were 66% more likely to be scored positively than identical sentences with "old" adjectives, across all sentiment analysis tools.
  • Implicit Age Bias: Sentences constructed with implicitly "young" adjectives were 1.09 times more likely to be scored positive, while implicitly "old" adjectives were 0.91 times as likely to be scored positive compared to control adjectives.
  • Bias Reduction: In Phase 3, the classifier trained on the "Age-Removed" corpus showed no significant bias (p<.7796), and the mean difference in confidence for "old" vs. "young" sentences being classified as positive was an order of magnitude lower (0.0063) compared to classifiers trained on data containing age-related tweets (0.0671 and 0.0669).

Limitations and Future Work

  • Model Architectures: The data-centric bias reduction method worked well for the tested Maximum Entropy classifiers but may not be directly applicable to other architectures like recurrent neural networks, which are more sensitive to word order and syntax.
  • Subtler Biases: The study acknowledges that its approach might not capture subtler forms of social bias, such as associations between broader topics and gender.
  • Intersectionality: The paper notes that social bias rarely exists along a single dimension, and future work should consider intersectional biases (e.g., age combined with gender or race).
  • Future Directions: The authors suggest exploring bias in relation to additional technical characteristics of models and further investigating the challenges of using computational techniques to study underrepresented groups. They also emphasize the importance of contextualizing model outputs and understanding their limitations, especially when analyzing data from specific populations.

Summary updated

BibTeX
@inproceedings{diaz2019addressing,
  author = {Mark Díaz and Isaac Johnson and Amanda Lazar and Anne Marie Piper and Darren Gergle},
  title = {{Addressing Age-Related Bias in Sentiment Analysis}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.24963/ijcai.2019/852},
}
Figure