Introduction
Intimacy is a fundamental social dimension of human interaction, yet it remains significantly under-researched in the field of Natural Language Processing (NLP). While existing research extensively covers sentiment analysis and offensiveness, the subtle nuances of how people share personal feelings or "self-disclose" in daily life are often overlooked. This paper introduces MINT, the first large-scale multilingual dataset designed to quantify intimacy in social media text, covering 13,372 tweets across 10 languages.
The challenge of quantifying intimacy is twofold. First, it is highly subjective; what one individual considers an intimate disclosure, another might find mundane. Second, there is significant cross-cultural variance, as intimacy is expressed and perceived differently across various languages and cultures. To address these challenges, the authors organized SemEval-2023 Task 9, a global competition where 45 teams competed to predict intimacy levels across diverse linguistic contexts.
The MINT Dataset and Methodology
The researchers developed a rigorous pipeline to create the MINT (Multilingual INTimacy) dataset. To ensure the dataset wasn't skewed toward "not intimate" content, they employed a "bucket sampling" technique, over-sampling tweets likely to contain higher levels of intimacy based on a preliminary model.
Each of the 13,372 tweets was rated by seven native speakers on a 5-point Likert scale (1: Not intimate at all, 5: Very intimate). Strict quality controls, including demographic balancing and attention checks, were implemented to ensure reliable human-aggregated scores.
The dataset includes training data for English, Spanish, French, Portuguese, Italian, and Chinese. Crucially, it also includes "zero-shot" test languages—Dutch, Korean, Hindi, and Arabic—to evaluate how well models generalize to languages they were not specifically trained on for this task.
Key Results and Performance
The competition saw 45 teams from over 20 countries. The top-performing team, LAZYBOB, achieved an overall Pearson’s $r$ (correlation) of 0.616. A key finding was that domain-specific pre-training is vital; the XLM-T model (pre-trained on Twitter data) consistently outperformed the standard XLM-RoBERTa, suggesting that understanding the "dialect" of social media (emojis, hashtags, informal grammar) is essential for social tasks.
However, the results also highlighted a significant performance gap between languages. While models performed exceptionally well on Spanish ($r \approx 0.78$) and English, they struggled with zero-shot predictions for culturally and linguistically distinct languages like Hindi ($r \approx 0.28$) and Korean ($r \approx 0.41$).
Insights and Future Directions
The performance disparity suggests that current multilingual models still struggle with "cross-lingual transfer" for subtle social nuances. The "Chinese Outlier" effect also showed that performance on Chinese tweets had lower correlation with Western European languages, indicating distinct cultural expressions of intimacy.
While most teams relied on traditional fine-tuning for regression, the authors suggest that future work should explore more sophisticated prompting strategies. Furthermore, because intimacy is inherently subjective, the authors plan to release individual labels and demographic data. This will allow researchers to move beyond mean scores and model how different groups—based on gender or age—perceive intimacy differently, ultimately leading to more socially aware dialogue systems and content moderation tools.
Summary updated