Introduction
Natural Language Processing (NLP) has successfully quantified various social dimensions such as power, status, and politeness. However, "intimacy"—a critical component of human connection—has remained difficult to measure at scale. Intimacy is not merely a matter of topic or self-disclosure; it is deeply encoded in the style and pragmatics of how we communicate.
Researchers have recently developed a computational framework to bridge this gap, focusing specifically on the intimacy of questions. By analyzing over 80 million questions across Reddit, Twitter, literature, and film, this study reveals how social norms, gender expectations, and social distance shape our linguistic choices.
Methodology
The study employed a multi-stage pipeline to create a reliable metric for linguistic intimacy:
- Annotation via Best-Worst-Scaling (BWS): To avoid the inconsistencies of traditional 1–10 scales, annotators compared sets of four questions to identify the most and least intimate. This produced a high-resolution ranking for 2,397 "gold standard" questions.
- Model Architecture: The researchers fine-tuned a RoBERTa model. It was first pre-trained on 3 million unannotated Reddit questions and then fine-tuned on the human-labeled dataset to predict continuous intimacy scores.
- Large-Scale Analysis: The model was applied to a massive corpus including 77.4 million Reddit entries, 1.04 million Twitter mentions, 2.02 million literary quotes from Project Gutenberg, and over 50,000 film dialogue snippets.
Social Distance and the "Stranger on a Train" Effect
A primary finding of the research is the quantitative confirmation of the "Stranger on a Train" effect. Social psychology has long hypothesized that intimacy follows a U-shaped curve relative to social distance. We are most intimate with those closest to us (where trust is high) and with total strangers (where the social risk of future consequences is low).
As shown in the network analysis of Twitter data, intimacy levels drop sharply when interacting with acquaintances (1–4 degrees of separation) before rising again for complete strangers (6+ degrees). This suggests that individuals modulate their language to protect social capital when the risk of judgment is highest.
Gender Dynamics and Internalized Norms
The study also explored how gender shapes intimacy. Across all datasets—including fictional works—male-to-male interactions consistently exhibited the lowest levels of intimacy.
Interestingly, this trend persisted in literature and film even when the authors were female. This suggests that gendered linguistic norms are not just personal preferences but are deeply ingrained cultural expectations. However, the data also indicated that the presence of a female audience can "license" men to use more intimate language, bringing their scores closer to those of female-to-female interactions.
Pragmatic Strategies for Navigating Risk
Beyond social context, the researchers identified specific linguistic tools used to manage intimacy:
- Hedging: As questions become more intimate, users increasingly employ hedging (e.g., "I might be wrong, but..."). This serves as a pragmatic safety net, reducing the pressure on the recipient to provide an equally intimate or potentially uncomfortable answer.
- Anonymity: On platforms like Reddit, users strategically utilize "throwaway" accounts to bypass social norms. Questions asked from these accounts were significantly more intimate than those from standard accounts, demonstrating that anonymity is an active tool for audience design.
Conclusion and Ethical Considerations
The fine-tuned RoBERTa model achieved a Pearson’s $r = 0.87$ with human judgments, proving that computational models can capture the subtle stylistic cues of intimacy. While this technology offers potential for improving empathetic AI and chatbots, the authors warn of "dual-use" risks. Such models could theoretically be used to harvest intimate data for malicious purposes. To mitigate this, the researchers suggest that platforms could use these metrics defensively, such as prompting users to consider anonymity when the model detects a highly intimate query.
Summary updated