The ``Colonial Impulse" of Natural Language Processing: An Audit of Bengali Sentiment Analysis Tools and Their Identity-based Biases

Title of the Paper

The “Colonial Impulse” of Natural Language Processing: An Audit of Bengali Sentiment Analysis Tools and Their Identity-based Biases

Paper Information

• Subject Area: Natural Language Processing (NLP), Algorithmic Bias, Postcolonial Studies, Human-Computer Interaction (HCI)
• Keywords: Algorithm Audit, Sentiment Analysis Tools, Bias, Identity, Coloniality

Research Background and Problem

• Identified Issues or Challenges:

  • NLP tools may reproduce colonial biases and values, yet there is a lack of research on the mechanisms through which these biases operate.
  • Significant inequality in language resources and research: for example, English has far more resources than Bengali, despite the two languages having comparable numbers of speakers.
  • Biases in non-English NLP systems and their impacts remain underexplored.

• Significance:

  • Sentiment analysis tools are widely used in practice, such as in content moderation and public opinion analysis, but identity biases within these tools may reinforce existing social injustices.
  • Given the complex interplay of colonial history, religion, gender, and ethnicity in South Asian Bengali culture, this region requires a specific research framework.

• Research Motivation:

  • To investigate whether and how existing Bengali sentiment analysis tools exhibit biases based on gender, religion, and ethnicity.
  • To understand how these biases relate to the demographic background of tool developers.

Solution

• Proposed Approach:

  • Conduct a systematic audit of Bengali sentiment analysis tools available on the Python Package Index (PyPI) and GitHub to examine identity-based biases.
  • Utilize the existing Bengali Identity Bias Evaluation Dataset (BIBED), which contains sentences capable of explicitly or implicitly expressing identity.

• Innovations:

  • Combine postcolonial studies with algorithm auditing to analyze how sentiment analysis tools reflect colonial social structures.
  • Introduce a dual-layer analysis method of implicit and explicit identity expressions to explore the tools' adaptability to complex social preferences.
  • Quantify biases across multiple dimensions (e.g., gender, religion, and ethnicity) within the same tool.

• Implementation Steps and Techniques:

  1. Tool Selection: Screened 13 sentiment analysis tools from PyPI and GitHub based on functionality and documentation completeness.
  2. Dataset Construction: Tested tool outputs using sentences from the BIBED dataset.
  3. Algorithm Audit: Input identical sentences (with variations in key identity terms) into different tools and statistically analyzed their sentiment scores.
  4. Statistical Analysis: Used Kruskal-Wallis and Mann-Whitney U tests to determine whether score differences were significant.
  5. Correlation Between Bias and Developer Background: Explored the relationship between published developer background information and the biases exhibited by the tools.

Research Findings

• Specific Findings:

  • Inconsistencies: Significant differences in sentiment scores for the same sentences across tools, indicating a lack of universality in sentiment measurement.
  • Differences in Implicit vs. Explicit Expressions: Sentences with explicit identity mentions (e.g., direct references to “Bangladesh”) were often assigned more negative sentiments, while implicit expressions using regional dialects scored higher.
  • Identity Bias:
    • 38% of tools favored female identities, while 30% favored male identities.
    • 30% of tools exhibited bias toward Hindu identities, while 38% favored Muslim identities.
    • 77% of tools were biased toward Bangladeshi identities, with only 15% favoring Indian identities.
  • Developer Background Correlation: While no direct relationship was observed between developer background and tool bias, the homogeneity of developers' ethnic and religious identities suggests potential limitations in development perspectives.

• Advantages Over Existing Solutions:

  • Conducted the first systematic algorithm audit of Bengali sentiment analysis tools.
  • Explored differences between explicit and implicit identity expressions in sentiment analysis for the first time.
  • Introduced the context of South Asian colonial history into the study of algorithmic bias, expanding the scope of bias research.

• Limitations and Future Directions:

  • Binary Identity Representation: This study only considered three binary classifications: male-female, Hindu-Muslim, and Bangladeshi-Indian, excluding broader groups (e.g., transgender individuals, other minority religions).
  • Intersectional Bias Analysis: Did not explore mechanisms of bias in intersectional identities (e.g., Muslim women, transgender Indians).
  • Future Work:
    • Use qualitative methods (e.g., interviews) to delve deeper into how tools handle more nuanced identity categories.
    • Investigate how different datasets contribute to the construction of tool biases.
    • Expand to other low-resource languages and analyze sampling biases related to other social dimensions (e.g., caste, sexual orientation).

This study highlights the limitations of Bengali sentiment analysis tools and their potential societal impacts while offering a method to incorporate postcolonial perspectives into technology research.