TL;DR
This paper tackles the pervasive issue of gender bias in word embeddings for Hindi and Marathi, two grammatically gendered languages. By developing novel evaluation metrics and debiasing techniques, the authors demonstrate a significant reduction in gender stereotypes within these embeddings, paving the way for fairer NLP applications in these widely spoken languages.
What Problem Does This Paper Address?
As natural language processing (NLP) systems become more integrated into our lives, the biases embedded within them become increasingly problematic. Word embeddings, the numerical representations of words that power many NLP tasks, are known to capture societal stereotypes. While much research has focused on debiasing English word embeddings, there's a significant gap in addressing this issue for low-resource languages, particularly those with grammatical gender. Languages like Hindi and Marathi assign a gender (masculine, feminine, and in Marathi, neutral) to every noun, which influences pronoun, adjective, and verb usage. This grammatical feature complicates the direct application of English-centric bias detection and mitigation methods. Without addressing this, NLP tools for these languages risk perpetuating harmful gender stereotypes in areas like occupation and emotion representation.
Methods
The authors propose a multi-step approach:
- Bias Quantification: They adapt existing metrics, the Embedding Coherence Test (ECT) and Relative Norm Distance (RND), to work with grammatically gendered languages.
- ECT: Measures the correlation between neutral word embeddings and gendered word embeddings. A higher correlation indicates less bias.
- RND: Captures the relative strength of association between neutral words and gendered groups. A lower value indicates less bias.
- They introduce specific versions:
ECT-nandRND-nfor neutral attributes (emotions, neutral occupations) andECT-gandRND-gfor gendered attributes (gendered occupations). - Gender Subspace Identification: To understand the direction of gender bias in the embeddings, they use Principal Component Analysis (PCA) and Relational Inner Product Association (RIPA) on pairs of masculine and feminine words.
- Bias Mitigation: Two debiasing techniques are explored:
- Projection: Making word vectors orthogonal to the identified gender direction.
- Partial Projection: A modified approach that projects out most of the gender direction but retains a small constant magnitude, aiming to preserve the inherent gender of words like "king" or "queen."
Datasets
The paper introduces new datasets for Hindi and Marathi:
- Occupation Datasets: Lists of 166 professions, categorized into gendered and neutral titles.
- Emotion Datasets: Words representing emotions, broadly classified into anger, fear, joy, and sadness.
- Seed Word Pairs: Pairs of masculine and feminine words (e.g., "king," "queen") used for identifying the gender subspace.
These datasets were verified by independent native speakers.
Key Insights and Ideas
- Grammatical Gender Matters: The presence of grammatical gender fundamentally changes how bias manifests and needs to be measured, requiring adaptations of existing metrics.
- Partial Projection is Promising: The "partial projection" debiasing method appears to be more effective than simple projection, especially for gendered words, as it avoids removing inherent gender information.
- Anger is Highly Biased: Across both Hindi and Marathi, the emotion "anger" shows the highest bias according to the ECT metric, suggesting a strong stereotypical association.
- Language-Specific Nuances: While both Hindi and Marathi exhibit bias, the paper's results tables (Tables 3 and 4) show varying degrees of effectiveness for debiasing techniques across languages and attribute types (neutral vs. gendered).
Key Results
The paper reports quantitative results in tables, showing the performance of debiasing techniques on various metrics. For instance:
- Hindi Emotions (Table 1): Partial projection with RIPA achieved the best ECT score (99.0%), and partial projection with PCA achieved the best RND score (0.3%).
- Hindi Occupations (Table 3): Partial projection with PCA was best for neutral occupations (ECT: 90.4%, RND: 1.79), while partial projection with PCA was also best for gendered occupations (ECT: 88.0%, RND: 1.70).
- Marathi Emotions (Table 2): Partial projection with PCA yielded the best ECT (96.1%) and RND (0.33) scores.
- Marathi Occupations (Table 4): Partial projection with RIPA showed the best results for gendered occupations (ECT: 64.2%, RND: 2.02).
In general, debiasing techniques significantly reduced bias, with partial projection often outperforming simple projection.
Limitations and Future Work
- Limitations: The authors acknowledge that their debiasing methods might alter some word vectors that are inherently gendered by definition. They hypothesize that partial projection performs better because it preserves some gender information.
- Future Work: The authors suggest extending these techniques to downstream NLP tasks to evaluate their impact on real-world applications. They also plan to experiment with debiasing contextual embeddings and large language models. Additionally, they aim to address other types of bias, such as religious, social, and cultural biases, which are particularly prevalent in Hindi and Marathi.
Summary updated