Focus

Social perception plays a central role in how language is interpreted: readers form impressions about intent, politeness, credibility, identity, and more from subtle linguistic cues.

These perceptions shape everyday interaction—how people respond, communicate, judge, trust, and engage—and they can also influence how language is used over time (e.g., style-shifting, spelling variation, and broader patterns of linguistic change).

Yet most NLP systems model these phenomena using surface-level proxies (e.g., fixed labels for “toxicity,” “politeness,” or “demographic information”), often treating socially grounded judgments as fixed properties of text. In doing so, they can collapse the distinction between what a text is intended to convey and how it is perceived in context. This limits our ability to build systems that reflect how humans interpret language across contexts and communities, and makes it difficult to understand what today’s large language models (LLMs) capture when they appear to make judgments about tone, credibility, harm, or social meaning.

NLPercep’26 aims to bridge this gap by bringing together researchers from NLP, computational social science, sociolinguistics, psychology, and related fields to study how language is perceived—and not just what it encodes.

The workshop places particular emphasis on the role of social perception in the era of large language models (LLMs) and evolving communication norms, and is grounded in three partially connected lines of work:

  • Sociolinguistics, where language attitudes and sociolinguistic perception examine how varieties, styles, and spellings are evaluated (e.g., “sounds professional,” “rude,” “educated”) and how these evaluations relate to linguistic variation and change.

  • Psychology and social perception, which studies how people form impressions, which cues drive judgments, and how context and norms shape perception and behavior.

  • Computational linguistics, where a growing body of work explicitly targets perceived constructs (e.g., perceived microaggression, perceived writing style, perceived gendered style, perceived racial bias), develops measurement and annotation schemes, and compares human and model judgments and impressions.

Questions we are interested in

We welcome theoretical, empirical, and computational contributions that address questions including (but not limited to):

  • Identity and demographic perception from text: What do humans infer about a speaker/author or a group from language (e.g., age, gender expression, regional background, social class, educational and cultural background) and what do human perceive from machine-generated texts (naturalness, trust, etc)?

  • LLMs as perceivers and social participants: Do LLM judgments align with human perception, where do they diverge, and what does that mean for real-world use—especially as LLMs increasingly mediate communication and shape emerging norms?

  • Self-perception and self-presentation in language: How do humans use language to position themselves (e.g., signaling uncertainty, confidence, competence with “this might be a dumb question, but…”)? Do LLMs display analogous self-perception (hedging, disclaimers, self-corrections, overconfidence), and how do these cues shape trust, interpretation, and interaction?

  • Cues that trigger social judgments: Which textual signals (linguistic and style signals) drive perceived intent, politeness/rudeness, credibility, warmth/competence, emotion (e.g., lexical choice, syntax, spelling variation, punctuation, emoji, hedging, register, code-switching, stance markers)?

  • Context and norm sensitivity: How do perceptions shift with context—topic, audience, relationship, platform, or community norms—and are LLMs sensitive to the same contextual cues?

  • Stereotypes, bias, and perception: When do humans and LLMs produce stereotype-consistent interpretations or judgments? How do such perceptions relate to social bias and stereotyping, and how can we study these effects from the perspective of perception?

  • Perception vs phenomenon: How should we distinguish constructs from their perception (e.g., microaggression vs perceived microaggression; hate speech vs perceived hate; gendered style vs perceived gendered style; emotion vs perceived emotion), and what does this imply for dataset design and evaluation, and model development?

  • Measuring social perception: What methodologies best capture perception reliably and validly (experimental designs, annotation schemes, disagreement-aware modeling)? How should we evaluate models when perceptions are variable, contextual, and norm-dependent?