💬 NLP

P1SCO: Social Dimensions from a Perspectivist Lens

The paper introduces P1SCO, a novel dataset of social media comments annotated across ten social dimensions with rich annotator metadata, designed to facilitate nuanced research into how social perceptions vary across platforms, individuals, and demographic factors.

Original authors: Amanda Cercas Curry, Gianmarco de Francisci Morales, Luca Maria Aiello

Published 2026-05-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Amanda Cercas Curry, Gianmarco de Francisci Morales, Luca Maria Aiello

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are looking at a single, strange painting hanging in a gallery. To you, it looks like a sad storm. To your friend next to you, it looks like a peaceful sunset. To a third person, it looks like a map of a city. None of you are "wrong." You are just looking at the same thing through different colored glasses.

This paper, P1SCO, is about building a massive library of those "colored glasses" to understand how people interpret social interactions online.

Here is the breakdown of what the researchers did, using simple analogies:

1. The Core Idea: Everyone Sees Things Differently

The authors start with a simple truth: Language isn't just about the words; it's about who is saying them and who is hearing them.

Think of a social media comment like a piece of music. If you play a jazz song for a classical music fan, a rock fan, and a jazz fan, they will all hear something different. The paper argues that when we read a comment on Instagram, Reddit, or YouTube, we don't just read the text; we filter it through our own life experiences, our personality, and our gender.

2. The "Ten Social Dimensions" (The Recipe Ingredients)

To study this, the researchers used a specific list of 10 "flavors" or social goals that people try to achieve when they talk. Think of these like the main ingredients in a soup:

Knowledge: Sharing info or teaching.
Power: Trying to control or command.
Status: Giving praise or showing respect.
Trust: Showing you can rely on someone.
Support: Being a friend or offering help.
Romance: Showing love or affection.
Similarity: Finding common ground.
Identity: Belonging to a group.
Fun: Making jokes or having a good time.
Conflict: Arguing or disagreeing.

3. The Experiment: A Giant Taste Test

The researchers didn't just ask people to read comments; they set up a massive taste test.

The Ingredients: They gathered 6,000 comments from three very different "restaurants" (social media platforms): Instagram (visual and personal), YouTube (video and broad), and Reddit (text-heavy and niche).
The Tasters: They recruited over 500 people from the US and UK. Before they started, they asked these tasters to fill out a questionnaire about their age, gender, politics, and even their personality type (using the famous "Big Five" personality test, which measures things like how open-minded or organized you are).
The Task: Each person read 50 comments and had to guess which "flavors" (from the list of 10 above) were present. They could pick more than one flavor for a single comment.

4. What They Found: The Glasses Matter

The most important discovery is that disagreement isn't a mistake; it's the point.

The Platform Effect: Just like a jazz club feels different from a rock concert, the platform changed how people saw things.
- On Reddit and YouTube, people mostly saw "Knowledge" and "Conflict" (debates and facts).
- On Instagram, people saw more "Support," "Similarity," and "Status" (praise and friendship).
- Analogy: It's like how a comment on a news site might look like a political argument, while the same comment on a photo of a cat might look like a friendly joke.
The Gender Effect: The study found that men and women often "taste" the same comment differently.
- Analogy: If a comment is ambiguous, men were more likely to taste "Power" or "Conflict," while women were more likely to taste "Support" or "Trust." It's as if men and women are wearing slightly different tinted glasses that highlight different parts of the conversation.
The Personality Effect: A person's personality also acted like a filter.
- People who were more "Open" to new ideas were more likely to spot "Knowledge."
- People who were more "Conscientious" (organized/dutiful) were more likely to spot "Support."
- Note: These effects were small, like a subtle spice in a soup, not a giant flavor explosion.

5. Why This Matters (According to the Paper)

Usually, when scientists collect data, they try to get everyone to agree on one "correct" answer. If 10 people look at a comment and 5 say it's "Funny" and 5 say it's "Rude," traditional science might throw that data away as "noise."

P1SCO says: Don't throw it away!
The disagreement is the data. It proves that social meaning is perspectival. The paper shows that who you are (your gender, your age, your personality) and where you are (which app you are on) fundamentally changes how you understand human interaction.

Summary

The paper built a dataset called P1SCO to prove that there is no single "objective" way to read a social media comment. Instead, every comment is interpreted through a unique lens made of the reader's identity and the platform they are on. By keeping all these different perspectives separate, rather than averaging them out, we get a much richer, more accurate picture of how humans actually connect.

Technical Summary: P1SCO – Social Dimensions from a Perspectivist Lens

Problem Statement
Current computational approaches to social pragmatics often treat linguistic meaning as objective or collapse diverse interpretations into a single "ground truth." However, language is inherently perspectival; social meaning is shaped by the interpreter's cultural background, personal experience, and identity. Previous datasets, such as the seminal work by Choi et al. (2020), are limited by their aggregation of labels (reporting only total counts per sentence), reliance on a single platform (Reddit), and lack of annotator metadata. This aggregation obscures the legitimate variation in social perception, treating annotator disagreement as noise rather than a meaningful signal of how different individuals interpret social intent. There is a need for a dataset that preserves individual annotator perspectives, captures cross-platform variations, and links social dimension labels to annotator demographics and personality traits.

Methodology
The authors introduce P1SCO (Perspectives on the 10 Social dimensions of Conversation), a dataset designed to operationalize a perspectivist approach to social pragmatics.

Data Collection: The dataset comprises 6,000 unique comments (2,000 each) from three distinct platforms: Instagram, YouTube, and Reddit. Comments were filtered to be between 20 and 100 words, non-emoji, and in English. The sampling strategy utilized the Instagram Influencer Dataset, the YouTube Search API with randomized ID generation, and the Pushshift Reddit dataset.
Annotation Framework: Unlike previous works that forced a single label, P1SCO allows annotators to assign multiple labels from a set of ten social dimensions defined by Choi et al. (2020): Knowledge, Power, Status, Trust, Support, Romance, Similarity, Identity, Fun, and Conflict. Annotators could also select "Other" (with specification) or "None."
Annotator Recruitment and Metadata: 543 participants were recruited via Prolific from the U.K. and U.S. To ensure data quality, the study verified demographic consistency with Prolific profiles, filtered out outliers based on completion time, and included duplicate examples for intra-annotator agreement checks.
Annotator Profiling: Before labeling, participants completed sociodemographic questionnaires and the 10-item Big Five personality inventory (Rammstedt and John, 2007). This resulted in a rich metadata set linking 28,775 annotations to specific annotator characteristics (gender, age, nationality, political affiliation, education, ethnicity, and personality traits).
Analysis: The authors employed mixed-effects logistic regression models to examine how annotator characteristics predict label assignment. Agreement was measured using pairwise Cohen's $\kappa$ (preferred over Krippendorff's $\alpha$ to capture localized agreement in perspectival tasks) and analyzed both overall and stratified by demographic groups and platforms.

Key Results

Label Distribution: The dataset is highly skewed. Knowledge (38.2%), Support (28.4%), and Similarity (25.7%) are the most frequent labels. Romance, Power, and Trust are comparatively rare. Most examples contain at least two overlapping dimensions, indicating that social interactions are rarely interpreted through a single lens.
Platform Effects: Platform context significantly influences perception. YouTube and Reddit show similar distributions dominated by Knowledge, Similarity, and Conflict. In contrast, Instagram interactions are characterized by Support, Similarity, and Status, with a notably higher prevalence of Romance and a lower proportion of "None" labels.
Demographic and Personality Influences:
- Gender: Male annotators assigned significantly more labels per example than females. While overall agreement was moderate, gendered differences in label selection were significant, particularly on Reddit and YouTube. The analysis revealed a perceptual analogue to Tannen's "genderlects": men were more likely to interpret interactions as Power or Conflict, while women were more likely to label the same interactions as Support, Trust, or Similarity.
- Personality: Correlations between Big Five traits and label assignment were generally weak but statistically significant. For instance, Openness showed a modest positive correlation with Knowledge annotations, while Conscientiousness was weakly associated with Support.
- Agreement: Homogeneous demographic groups (e.g., same gender or nationality) exhibited higher within-group agreement than the general population, suggesting shared experiences create convergent interpretive frameworks.
Modeling: Mixed-effects models confirmed that the source platform is the strongest predictor of label assignment, followed by sociodemographic factors (particularly gender, age, and education).

Significance and Contributions
The paper claims three primary contributions:

First Perspectivist Dataset: P1SCO is the first dataset to provide unaggregated, multi-label annotations of social dimensions linked to rich annotator metadata, moving beyond the "ground truth" paradigm.
Empirical Evidence for Perspectivism: The study demonstrates that demographic characteristics and personality traits systematically shape how social dimensions are perceived and labeled. Disagreement is reframed not as error, but as a reflection of genuine diversity in social perception.
Methodological Guidance: The authors provide a framework for perspectivist annotation that explicitly preserves individual perspectives, models annotator characteristics, and transparently reports demographic composition.

Limitations
The authors acknowledge several constraints:

Interpretive Nature: Labels reflect annotator perception rather than objective text properties, limiting their use as definitive ground truth for classification.
Cognitive Load: The task of identifying ten dimensions is cognitively demanding, and the lack of explicit rationales limits insight into the reasoning behind disagreements.
Diversity: The annotator pool is limited to the U.S. and U.K. and lacks ethnic diversity, bounding the dataset's cultural scope.
Generalizability: Findings are specific to the three sampled platforms and may not extend to other social media contexts.
Label Skew: The rarity of certain dimensions (e.g., Romance, Power) creates skewed distributions that may challenge the training of balanced supervised models without resampling.

Conclusion
P1SCO establishes that social meaning is not uniform but is enacted and interpreted through personal and situational filters. By preserving annotator-level variation, the dataset enables the study of how social goals are communicated and understood across diverse individuals, supporting future research into inter- and intra-annotator agreement and the influence of identity on social discourse.