Silver-Tongued and Sundry: Exploring Intersectional Pronouns with ChatGPT

Document Title

Silver-Tongued and Sundry: Exploring Intersectional Pronouns with ChatGPT

Document Information

• Subject Area: Human-Computer Interaction (HCI), Natural Language Processing, Social Identity Perception
• Keywords: Conversational User Interface, Chatbot, Gender, Japan, First-Person Pronouns, Intersectionality, Social Identity, ChatGPT

Research Background and Problem

• Identified Problem: Current conversational systems built on large language models (LLMs), such as ChatGPT, possess near-human interaction capabilities. However, these systems may (unintentionally) simulate human social identities, reflecting biases or societal stereotypes.
• Significance: Designing intelligent assistants or interactive agents that include explicit social identity cues (e.g., pronouns) could reinforce or challenge societal biases, raising ethical concerns, particularly in contexts where diverse social identities intersect.
• Research Motivation: In Japanese language usage, first-person pronouns are strong markers of social identity, reflecting not only gender but also age, region, class, and formality. It remains unclear whether chatbots can evoke user perceptions of these social identities through pronoun usage.
• Related Work: Most existing research on identity perception focuses on single-dimensional identity categories (e.g., gender) and predominantly targets English-speaking contexts, with limited attention to multidimensional intersectional identities and non-Western settings.

Solution

• Proposed Method: Design an online experiment where ChatGPT interacts with users using 10 different Japanese first-person pronouns (intersectional pronouns, kousa-daimeishi) to analyze Japanese users' perceptions of ChatGPT’s gender, age, region, and formality.
• Innovations:
  • Explores the potential for intelligent agents to simulate intersectional social identities through linguistic pronouns.
  • Provides a culturally sensitive, language-based methodology for designing intelligent agents.
  • Compares user perception differences between participants from different regions (Kanto and Kansai).
• Implementation Steps:
  1. Participant Recruitment: Recruit 201 participants from Japan's Kanto and Kansai regions via an online platform (Yahoo! Crowdsourcing).
  2. Experiment Design: ChatGPT uses 10 different Japanese first-person pronouns in video-based interactions with users.
  3. Data Collection: Gather quantitative ratings and qualitative evaluations of ChatGPT’s perceived gender, age, similarity, and other identity characteristics from participants.
  4. Data Analysis: Use qualitative and quantitative methods to analyze the impact of pronoun usage on identity perception.

Research Outcomes

• Specific Findings:
  • First-person pronouns alone can evoke user perceptions of both simple (e.g., gender) and complex (e.g., intersectional) social identities in ChatGPT.
  • 私 (watashi-kanji, considered a neutral pronoun) was perceived as gender-ambiguous by some users but was more commonly categorized as feminine.
  • Pronouns showed significant differences in perceived gender, age, formality, and regional associations. Key findings include:
    1. Gender Perception: わたし (watashi) and あたし (atashi) were perceived as more feminine, while ぼく (boku) and おれ (ore) were seen as masculine pronouns.
    2. Age and Formality Perception: わし (washi) was associated with older age, informality, and rural backgrounds, while あたくし (atakushi) was perceived as feminine and highly formal.
    3. Regional Differences: Although not entirely as expected, subtle differences in perceptions of certain pronouns (e.g., うち (uchi)) were observed between Kanto and Kansai participants.
• Advantages Compared to Existing Solutions:
  • Proposes a rapid method for constructing intersectional social identity personas, particularly in Japanese-language interaction scenarios.
  • Goes beyond gender to analyze dimensions such as age and regional identity.
• Experimental or Evaluation Results:
  • Quantitative analysis (e.g., Chi-square tests) confirmed the significance of the observed perceptions.
  • Qualitative analysis highlighted regional or formality-related associations triggered by different pronouns.
• Limitations and Future Directions:
  • The gender distribution of participants was uneven, with more female participants, potentially affecting the representativeness of results.
  • Broader linguistic features (e.g., sentence-ending particles) were not analyzed.
  • Future research could expand the regional scope, explore the relationship between pronouns and context (e.g., conversation scenarios, user tasks), and investigate how other languages reflect intersectional identities.