Potential Pitfalls With Automatic Sentiment Analysis: The Example of Queerphobic Bias

Area of Science:

• Natural Language Processing (NLP)
• Artificial Intelligence (AI)
• Sociolinguistics

Background:

• Automated sentiment analysis is widely used for trend detection across various domains.
• Natural Language Processing (NLP) techniques, including sentiment analysis, can inadvertently perpetuate societal biases.
• Existing research has identified bias in sentiment analysis concerning gender, ethnicity, and disability.

Purpose of the Study:

• To investigate bias in popular sentiment analysis tools specifically concerning queer identities.
• To expand upon existing research by examining a broader range of marginalized groups within sentiment analysis.
• To provide guidance on selecting sentiment analysis tools to mitigate bias.

Main Methods:

• Evaluation of six popular sentiment analysis tools.
• Testing tools with sentences related to various queer identities.
• Comparative analysis of tool responses to identify biased outputs.

Main Results:

• Evidence of bias against several marginalized queer identities was found across tested tools.
• Two prominent models (Google, Amazon) showed bias despite apparent superficial debiasing efforts.
• The extent and nature of bias varied among the evaluated sentiment analysis tools.

Conclusions:

• Sentiment analysis tools are not universally unbiased and can disadvantage marginalized communities.
• Superficial debiasing methods may be insufficient to eliminate bias in AI models.
• Selecting sentiment analysis tools requires careful consideration of potential biases to ensure reliable and equitable results.