Quantifying Intimacy in Language

Jiaxin Pei, David Jurgens

Abstract

This paper presents a computational framework for quantifying intimacy in language, focusing on the pragmatic moves within questions. By training a RoBERTa-based model on human-labeled data using Best-Worst Scaling, the authors analyze 80.5 million questions across Reddit, Twitter, literature, and film. The study provides empirical evidence for social psychology theories regarding gender norms, social distance, and the strategic use of anonymity.

Introduction

While NLP has advanced in detecting topic and sentiment, the relational dimension of intimacy remains under-explored. Intimacy—the degree of warmth and willingness to share personal information—is contextual and subtle, often encoded in linguistic choices like hedging and swearing. This research focuses on questions as fundamental units of interpersonal exchange that require speakers to calibrate their intimacy levels to their audience.

Methodology

The researchers developed an intimacy dataset of 2,397 questions annotated via Best-Worst Scaling (BWS). They fine-tuned a RoBERTa-base model to predict intimacy scores, comparing it against baselines like Bag-of-Words and LDA Topic Models. This model was then applied to a massive dataset of 80.5 million questions from diverse sources, including social media and literature.

Key Results

The fine-tuned RoBERTa model achieved high performance, demonstrating that intimacy is deeply embedded in syntax and phrasing rather than just vocabulary.

Model performance comparison across different architectures.

The model reached a Pearson’s $r = 0.87$, significantly outperforming the Topic Model baseline ($r = 0.62$). It also showed robustness across domains, maintaining strong correlations when applied to literature and film data despite being trained primarily on Reddit.

Key Insights and Social Dynamics

The large-scale analysis revealed several key patterns in how intimacy is negotiated in social contexts:

The "Strangers on a Train" Effect

The data confirms a U-shaped relationship between intimacy and social distance. Individuals are most intimate with close friends (low risk) and total strangers (low social consequence), while being least intimate with casual acquaintances where the social cost of oversharing is highest.

The "Strangers on a Train" effect: Intimacy vs. Degrees of Separation in a network.

Gender Norms and Anonymity

  • Internalized Gender Norms: Male-to-male interactions showed the lowest intimacy levels across all datasets, including fiction. This suggests that hegemonic norms of masculinity are deeply ingrained in both real and imagined interactions.
  • Strategic Anonymity: Reddit users with "throwaway" accounts ask significantly more intimate questions than standard accounts, suggesting anonymity is a deliberate strategy to bypass social norms.
  • Linguistic Buffering: As questions become more intimate, speakers are more likely to use hedging (e.g., "might," "possibly") to reduce the risk of social awkwardness.

Discussion and Limitations

The study is limited by its binary focus on gender and its scope on questions rather than full conversations. Future work should include non-binary perspectives and broader linguistic contexts. Ethically, the authors warn against "intimacy mining" and suggest using these models to enhance user privacy on social platforms by recommending anonymous accounts for highly intimate queries.

Summary updated

Abstract

Intimacy is a fundamental aspect of how we relate to others in social settings. Language encodes the social information of intimacy through both topics and other more subtle cues (such as linguistic hedging and swearing). Here, we introduce a new computational framework for studying expressions of the intimacy in language with an accompanying dataset and deep learning model for accurately predicting the intimacy level of questions (Pearson's r=0.87). Through analyzing a dataset of 80.5M questions across social media, books, and films, we show that individuals employ interpersonal pragmatic moves in their language to align their intimacy with social settings. Then, in three studies, we further demonstrate how individuals modulate their intimacy to match social norms around gender, social distance, and audience, each validating key findings from studies in social psychology. Our work demonstrates that intimacy is a pervasive and impactful social dimension of language.
BibTeX
@inproceedings{pei2020quantifying,
  author = {Jiaxin Pei and David Jurgens},
  title = {{Quantifying Intimacy in Language}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2020},
  doi = {10.18653/v1/2020.emnlp-main.428},
}
Figure