Large Language Model Interview Response Gender Bias Dataset: A Bilingual Comparative Study Between English and Italian

The issue of AI fairness has received much attention, especially the bias in content generated by large language models (LLMs). The open-source Thesis_Dataset focuses on interview scenarios, comparing gender bias patterns in English and Italian LLMs, revealing potential stereotypes, and is of great value for understanding and improving AI fairness. Through rigorous design, this study provides a scenario-based, cross-lingual empirical foundation for AI fairness research, and open-sources the dataset to promote academic accumulation.

Large language models are widely used in the human resources field (resume screening, interview question generation, etc.). If they carry gender bias, it may exacerbate workplace inequality. Historically, technical systems have exposed gender bias multiple times. Due to the large scale and diverse sources of training data, LLMs may encode hidden and complex biases. As a key link in job hunting, gender differences in AI-generated content for interviews directly affect career opportunities and require focused attention.

The dataset selects English (no grammatical gender) and Italian (strict grammatical gender) for comparison, reflecting differences in language structure and cultural background:

• Language type differences: English has no grammatical gender except for pronouns, while Italian nouns, adjectives, etc., require gender changes, which can reveal the impact of language structure on bias;
• Cultural background differences: The UK and Italy represent different gender concepts and workplace cultures, distinguishing between universal patterns and specific cultural reflections;
• Data collection: Control variables such as job descriptions and qualifications to isolate the impact of gender factors on model output.

Gender bias detection covers:

• Stereotype reproduction: Association between occupations/traits and gender (e.g., leadership vocabulary with males, care-related vocabulary with females);
• Differences in ability assessment: Differences in evaluation language for different genders with the same qualifications (certainty, degree of praise);
• Differences in opportunity recommendation: Gender differences in career development path recommendations;
• Differences in language style: Tone certainty, detail level of suggestions, etc. Potential findings include: confirmation of the existence of bias, language adjustment effects, model differences, impact of prompt engineering, etc.

Research Contributions:

• Scenario focus: The high-impact interview scenario is close to practical applications;
• Cross-language perspective: Enriches the dimensions of bias research;
• Open-source sharing: Promotes academic reproduction and expansion;
• Methodological reference: Provides reference for similar studies. Social Significance:
• Algorithm accountability: Provides quantitative detection tools;
• Policy formulation: Supports AI fairness standards and audit mechanisms;
• Public awareness: Enhances understanding of AI ethics;
• Industry practice: Provides risk warnings for enterprise recruitment AI applications.

Mitigation Strategies:

• Data level: Balance bias in training data;
• Model level: Introduce fairness constraints (adversarial debiasing, regularization);
• Inference level: Debiasing prompts, output filtering;
• Evaluation level: Improve bias evaluation benchmarks. Limitations: Limited scenario representativeness, gender binary framework, few languages covered, static snapshot. Future Directions: Cover more languages and cultures, explore intersectionality (interaction between gender and race, etc.), verify mitigation technologies, and establish continuous monitoring mechanisms.