Scene Abstraction for Lexical Semantics: Structured Representations of Situated Meaning

2026-05-21 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

Scene Abstraction is a novel framework designed to create structured representations of the interpretive scenes associated with words across various usage contexts. This framework addresses the implicit situated dimensions of lexical meaning, such as the distinct situations and affective associations evoked by words like "coffee" versus "tea." Each scene is composed of a Contextual Scene, detailing Events, Entities, and Setting, and an expression-centered Expression Profile, which includes Engaged events, Generalizable properties, and Evoked emotions. The framework is operationalized through few-shot prompting of a large language model. Key contributions include the structured representation framework itself, the COCA-Scenes dataset comprising 520 usage instances for 26 keywords, and empirical validation. Experiments show scenes are reliably identifiable by human observers with 82.4% accuracy, an 11.8 percentage point improvement over text-only embeddings, and that scene profiles align more closely with human interpretation than ATOMIC-based alternatives, demonstrating an 86.4% preference across three semantic dimensions.

Key takeaway

For NLP Engineers developing advanced semantic models, consider integrating "Scene Abstraction" to capture the situated and affective dimensions of word meaning. Your current text-only embeddings may miss crucial contextual nuances, as evidenced by the 11.8 percentage point accuracy gain over traditional methods. By adopting structured scene representations, you can achieve better alignment with human interpretation, improving model performance in tasks requiring deep contextual understanding. Explore few-shot prompting with large language models to operationalize this framework effectively.

Key insights

"Scene Abstraction" provides a structured framework to represent the situated, interpretive dimensions of lexical meaning using LLMs.

Principles

• Lexical meaning extends beyond explicit properties to situated contexts.
• Structured scene representations enhance human interpretation alignment.

Method

The framework operationalizes scene identification by defining Contextual Scenes (Events, Entities, Setting) and Expression Profiles (Engaged events, Generalizable properties, Evoked emotions), then uses few-shot prompting of a large language model.

In practice

• Develop datasets like COCA-Scenes for situated meaning analysis.
• Evaluate scene representations against human interpretation benchmarks.

Topics

• Lexical Semantics
• Scene Abstraction
• Situated Meaning
• Large Language Models
• Few-shot Prompting
• COCA-Scenes Dataset

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Probing Semantic Alignment, Lexical Invariance, and Syntactic Influence in LLM Metaphor Processing — LLMs struggle with deep metaphor comprehension, often misinterpreting concepts and relying on lexical associations over context or full syntactic understanding.
• Geometry of Semantic Space: Comparative Study of Discrete and Continuous Models — Graph-based semantic models offer more interpretable and structured meaning representations than high-dimensional transformer embeddings.
• Learning Context-Conditioned Predicate Semantics via Prototype Feedback — AlignG dynamically adapts predicate semantics to image context using prototype feedback, improving scene graph generation accuracy.
• What They Saw, Not Just Where They Looked: Semantic Scanpath Similarity via VLMs and NLP metric — VLMs and NLP metrics can quantify semantic similarity in eye-tracking scanpaths, complementing traditional spatial analysis.
• Token Embeddings : From numbers to meaning — Token embeddings convert arbitrary token IDs into dense, learned vectors that capture semantic relationships through contextual co-occurrence.
• Hybrid Human-LLM Corpus Construction and LLM Evaluation for the Caused-Motion Construction — LLMs struggle with complex linguistic constructions like CMC, necessitating novel corpus creation methods.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.