Scene Abstraction for Lexical Semantics: Structured Representations of Situated Meaning

# Structured Representations of Situated Meaning
Source: [https://arxiv.org/html/2605.22542](https://arxiv.org/html/2605.22542)
## Scene Abstraction for Lexical Semantics: Structured Representations of Situated Meaning

Katrin Erk University of Massachusetts, Amherst Department of Linguistics kerk@umass\.edu

###### Abstract

Coffeeandteashare many properties, yet they evoke strikingly different situations, atmospheres, and affective associations\. These situated dimensions of word meaning are real and systematic, but they remain implicit in most computational representations of lexical meaning\. We proposeScene Abstraction, a framework for constructing structured representations of the interpretive scenes that words participate in across usage contexts\. Each scene consists of aContextual Scene\(Events, Entities, Setting\) and an expression\-centeredExpression Profile\(Engaged events, Generalizable properties, Evoked emotions\), operationalized through few\-shot prompting of a large language model\. Our contributions are threefold: \(1\) a structured representation framework for situated lexical meaning; \(2\) COCA\-Scenes, a dataset of 520 usage instances across 26 keywords for distinct scene identification; and \(3\) empirical evidence from two experiments suggesting that scenes are reliably identifiable across human observers \(82\.4% accuracy, \+11\.8 pp over text\-only embeddings\) and that our scene profiles more closely align with human interpretation of words in context than ATOMIC\-based alternatives \(86\.4% preference across three semantic dimensions\)\.

Scene Abstraction for Lexical Semantics: Structured Representations of Situated Meaning

Yejin ChoThe University of Texas at AustinDepartment of Linguisticsycho@utexas\.eduKatrin ErkUniversity of Massachusetts, AmherstDepartment of Linguisticskerk@umass\.edu

Contextual scene𝒞\\mathcal\{C\}Expression profileℰ\\mathcal\{E\}\(coffee\)Input text“She drank her thirdcoffeeof the morning and kept typing\.”Scene abstractiongpt\-4o\-miniEventsPersonXdrinksObjectYPersonXtypes continuouslyEntitiesPersonX\(she\): productive, alertObjectY\(coffee\):stimulating, caffeinatedSettingTime: morning;Place: unspecified;Atmos\.: focused, energeticEngaged events\-PersonXdrinks it\- SupportsPersonX’s workGeneralizable properties\- Boosts productivity\- Work ritual for energyEvoked emotions\- Motivation\(The act of drinking coffee suggests a drive to continue working\.\)𝒮​\(u,x\)=⟨𝒞,ℰ⟩\\mathcal\{S\}\(u,x\)=\\langle\\,\\mathcal\{C\},\\;\\mathcal\{E\}\\,\\rangle

Figure 1:The Scene Abstraction framework\. Given a usage contextuand target expressionx\(coffee\) in it, an LLM produces𝒮​\(u,x\)\\mathcal\{S\}\(u,x\)comprising aContextual Scene𝒞\\mathcal\{C\}\(Events, Entities, Setting\) that captures the overall situation described byu, and anExpression Profileℰ\\mathcal\{E\}\(Engaged events, Generalizable properties, Evoked emotions\) that characterizes the scene\-grounded meaning ofx\.## 1Introduction

What doescrowmean? The dictionary answer—a large black bird—captures little of how the word actually functions in context\. In practice,crowparticipates in scenes of silence, isolation, ominous atmospheres, winter landscapes, and death symbolism; these recurring situational patterns are part of what it means to know the word, yet none are explicitly encoded in its dictionary definition\. Similarly, words such aswhiskey,rain, orroseevoke recurring experiential and affective situations that extend well beyond referential content\. These situated dimensions of meaning are not incidental; they reflect structured interpretive regularities that shape how words are understood in context\.

Capturing this situated, interpretive character of meaning computationally, however, remains an open challenge\. Frame semantics\(Fillmore,[1976](https://arxiv.org/html/2605.22542#bib.bib6),[1982](https://arxiv.org/html/2605.22542#bib.bib7)\)proposes that word meaning is grounded in rich background knowledge, including situational, affective, and cultural dimensions\. The primary large\-scale implementation of this tradition, FrameNet\(Bakeret al\.,[1998](https://arxiv.org/html/2605.22542#bib.bib8)\), however, focuses on a narrower set of semantic frames describing predicate\-argument structure\. It does not readily scale to the open\-ended, contextually variable dimensions of situated meaning, such ascrowbeing ominous in one scene and social or intelligent in another\. Distributional models\(Harris,[1954](https://arxiv.org/html/2605.22542#bib.bib1); Landauer and Dumais,[1997](https://arxiv.org/html/2605.22542#bib.bib2)\)and their contextual successors\(Peterset al\.,[2018](https://arxiv.org/html/2605.22542#bib.bib3); Devlinet al\.,[2019](https://arxiv.org/html/2605.22542#bib.bib4)\)capture rich usage\-based patterns and scale to large corpora, but encode semantic structure—events, participant roles, atmosphere, affective associations—implicitly within dense vectors that are difficult to inspect, compare, or interpret directly\. Instruction\-tuned LLMs demonstrate strong contextual understanding and can produce nuanced interpretations of words in context, but their default outputs are free\-form prose—unstructured, variable across prompting conditions, and difficult to compare or aggregate systematically across usage instances\. The framework we propose draws on all three traditions, operationalizing the situated, affective, and cultural dimensions of word meaning that frame semantics foregrounds, with the scalability of distributional approaches and the generative capacity of LLMs\.

We proposeScene Abstraction, a framework that views lexical meaning as structured distributions over interpretive scenes𝒮​\(u,x\)\\mathcal\{S\}\(u,x\), whereuuis a usage context andxxis a target expression\. The central representational idea is that some aspect of word meaning in context can be more effectively captured by the structured scene it instantiates than by its surface textual surroundings alone\. This framework attempts to understand aspects of lexical meaning through the characteristic patterns of scenes across usage instances of a word\. As illustrated in Figure[1](https://arxiv.org/html/2605.22542#S0.F1), each scene consists of two complementary components: aContextual Scene𝒞\\mathcal\{C\}capturing the broader situational interpretation—events, entities, and setting—and, crucially, a target\-expression\-centeredExpression Profileℰ\\mathcal\{E\}capturing the scene\-grounded meaning ofxxspecifically: the events it engages in, the properties it bears within the scene, and the affective associations it evokes\. Different words in the same sentence will have different Expression Profiles: it is the Expression Profile that makes this per\-expression meaning realization explicit and inspectable\. Scenes are not predefined ontological categories but open\-ended interpretive representations constructed dynamically for each usage instance\.

We evaluate the framework through two experiments addressing distinct validation goals\. First, anodd\-scene\-outtask assesses construct validity: do scene\-level distinctions constitute a reliably shared discriminative structure among humans, and do automatically generated scene\-based representations outperform raw\-text embeddings in capturing them? Second, a human preference study evaluates output quality: do LLM\-generated scene abstractions align with human interpretation of words in context? Specifically, we compare them against ATOMIC\-based commonsense profiles\(Sapet al\.,[2019](https://arxiv.org/html/2605.22542#bib.bib16); Hwanget al\.,[2021](https://arxiv.org/html/2605.22542#bib.bib18)\), which represent the closest existing alternative for structured situational inference\.

This paper makes the following contributions:

- •We introduceScene Abstraction, a two\-layer framework comprising a Contextual Scene𝒞\\mathcal\{C\}\(events, entities, setting\) and a target\-expression\-centered Expression Profileℰ\\mathcal\{E\}, operationalized via few\-shot prompting of a large language model\.
- •We constructCOCA\-Scenes, a dataset of 520 usage instances across 26 keywords designed to support scene\-level evaluation\.
- •We provide empirical evidence that scene\-based representations outperform text\-only embeddings in human\-aligned situational discrimination \(69\.3% vs\. 57\.5% accuracy\), and that LLM\-generated scene abstractions align substantially more closely with human interpretation of words in context than ATOMIC\-base alternatives \(86\.4% preference\)\.

## 2Related Work

Existing approaches to lexical meaning representation, characterization of situations, and computational frameworks for commonsense knowledge offer important insights on which the current work builds\.

##### Frame Semantics\.

Frame Semantics\(Fillmore,[1976](https://arxiv.org/html/2605.22542#bib.bib6),[1982](https://arxiv.org/html/2605.22542#bib.bib7)\)proposed that understanding a word requires access to rich background knowledge about the situations, participants, and relations it evokes\. Construction grammar has similarly argued that meaning is inseparable from structured situational knowledge\(Goldberg,[1995](https://arxiv.org/html/2605.22542#bib.bib9)\)\. Scene Abstraction builds on this foundational insight, extending it computationally through the use of LLMs to construct open\-ended, usage\-grounded scene representations for any expression and context\. The primary large\-scale implementation of this tradition, FrameNet\(Bakeret al\.,[1998](https://arxiv.org/html/2605.22542#bib.bib8)\), focuses on a particular facet of Frame Semantics: semantic frames describing the core meaning and predicate\-argument structure\. Scene Abstraction, by contrast, requires no predefined inventory for interpretation, allowing it to flexibly capture how the same word takes on different meanings depending on context, such ascrowbeing ominous in one scene and intelligent or social in another\.

##### Situation Models, Scripts, and Scene\.

Research on discourse comprehension has established that language understanding involves constructing rich mental representations beyond surface linguistic form\(Van Dijket al\.,[1983](https://arxiv.org/html/2605.22542#bib.bib10); Zwaanet al\.,[1995](https://arxiv.org/html/2605.22542#bib.bib12); Zwaan and Radvansky,[1998](https://arxiv.org/html/2605.22542#bib.bib11)\)\. Script\-based approaches further emphasize stereotyped event sequences in organizing situational knowledge\(Schank and Abelson,[1977](https://arxiv.org/html/2605.22542#bib.bib13)\), and such structures have been induced from large text corpora\(Chambers and Jurafsky,[2008](https://arxiv.org/html/2605.22542#bib.bib14); Pichotta and Mooney,[2016](https://arxiv.org/html/2605.22542#bib.bib15)\)\. Scene Abstraction draws on this tradition but differs in two respects\. Unlike situation models, which remain implicit cognitive representations, scene abstractionexternalizesthem in explicit, structured, and computationally accessible form\. Unlike scripts, which are organized around temporal event sequences, a scene is closer to asituational snapshot: the structured interpretation formed at a given moment, in which events, participants, atmosphere, and affect are jointly organized within a single instance\. The scene abstraction approach prioritizes this within\-moment interpretive variability over the sequential structure that scripts encode\.

##### Distributional and Contextual Semantics\.

Distributional and contextual embedding models\(Harris,[1954](https://arxiv.org/html/2605.22542#bib.bib1); Lee and Goldwasser,[2018](https://arxiv.org/html/2605.22542#bib.bib29); Peterset al\.,[2018](https://arxiv.org/html/2605.22542#bib.bib3); Devlinet al\.,[2019](https://arxiv.org/html/2605.22542#bib.bib4); Wiedemannet al\.,[2019](https://arxiv.org/html/2605.22542#bib.bib24); Schlechtweget al\.,[2020](https://arxiv.org/html/2605.22542#bib.bib23)\)offer strong computability, but encode semantic structure implicitly within dense vectors that do not directly expose the event structures, participant relations, atmosphere, or affective dimensions important for situated lexical interpretation\.Hoyleet al\.\([2023](https://arxiv.org/html/2605.22542#bib.bib30)\)show that externalizing implicit communicative content via LLM\-generated propositions yields better text representations across downstream tasks\. Scene Abstraction is motivated by a similar observation but targets specific situated dimensions of lexical meaning—events, participants, atmosphere, and expression\-specific profiles—organized into a fixed, reusable schema for cross\-instance comparison\. Because each scene component can be independently encoded as a text embedding, the resulting distributions are both interpretable and computable\.

##### Commonsense and Event Knowledge Representation\.

Large\-scale commonsense knowledge graphs\(Sapet al\.,[2019](https://arxiv.org/html/2605.22542#bib.bib16); Bosselutet al\.,[2019](https://arxiv.org/html/2605.22542#bib.bib17); Hwanget al\.,[2021](https://arxiv.org/html/2605.22542#bib.bib18)\)and narrative\-grounded causal knowledge resources\(Mostafazadehet al\.,[2020](https://arxiv.org/html/2605.22542#bib.bib33)\)have substantially advanced the representation of implicit world knowledge—causal relations, social dynamics, and event structure\. These resources are primarily designed to support general\-purpose commonsense inference and narrative understanding, however, rather than the representation of lexical meaning specifically\. Their relational schemas describe what typically holds in everyday situations in general, and do not directly address how the meaning of a particular word is shaped by the specific context in which it appears\. Scene Abstraction shares the goal of making implicit situational knowledge explicit, but focuses specifically on what a target expression does, implies, and evokes in a given usage instance—a question that commonsense knowledge resources are not designed to answer\. This distinction is directly tested in Experiment 2, where ATOMIC\-based profiles consistently produce outputs less aligned with human scene\-grounded interpretation of target words than Scene Profiles\.

##### LLMs for Structured Semantic Generation\.

LLMs have demonstrated strong capabilities for generating structured semantic representations\(Brownet al\.,[2020](https://arxiv.org/html/2605.22542#bib.bib5); Weiet al\.,[2022](https://arxiv.org/html/2605.22542#bib.bib20); Westet al\.,[2022](https://arxiv.org/html/2605.22542#bib.bib19); Gilardiet al\.,[2023](https://arxiv.org/html/2605.22542#bib.bib21); Wanget al\.,[2024](https://arxiv.org/html/2605.22542#bib.bib22)\)\. Closely related to our work,Guet al\.\([2022](https://arxiv.org/html/2605.22542#bib.bib31)\)shows that explicitly elaborating situational context before answering improves downstream QA performance, providing evidence that structured situational elaboration is a productive intermediary step for language understanding\. Scene Abstraction builds on these capabilities by supplying a fixed representational schema that channels LLM interpretive capacity into more structured and explicit scene representations\.

Scene Abstraction inherits from Frame Semantics the core idea that word meaning is grounded in rich background knowledge, including situational, affective, and cultural dimensions\. It draws further on the situational grounding of discourse comprehension research and the scalability of distributional approaches\. What it contributes is a computational method for operationalizing this background knowledge at the level of individual usage instances: based on a manually designed scene schema, we use LLMs to construct structured, easily interpreted scene representations for any word in any context with great flexibility\.

## 3Scene Abstraction Framework

### 3\.1Motivation

A central observation motivating our framework is that word meaning in context is not simply a function of local lexical co\-occurrence\. When readers encounter a sentence such as:

> The man sat alone at the kitchen table, drinking whiskey late at night\.

their interpretation ofwhiskeydraws on a structured understanding of the entire situation: the solitude, the hour, the setting, the emotional weight implied by the scene\. This interpretive structure—events, participants, atmosphere, and the target expression’s role within it—is central to how the word is understood in context\. A man drinking whiskey alone at night and a man drinking beer with friends at a pub involve the same surface event—drinking an alcoholic beverage—yet give rise to substantially different scenes\. Distinguishing these cases may require a representation that encodes more than just the surface event\.

We formalize this observation through the Scene Abstraction framework: the contextual meaning ofxxinuucan be more effectively captured by a structured scene representation than by surface textual context alone, and type\-level lexical meaning can be understood as a distribution of such scenes across usage instances\.

Table 1:Scene output \(abbreviated\) forwhiskeyin“The man sat alone at the kitchen table, drinking whiskey late at night\.”, generated bygpt\-4o\-mini\. The Contextual Scene𝒞\\mathcal\{C\}captures the broader situational interpretation; the Expression Profileℰ\\mathcal\{E\}captures the scene\-grounded meaning of the target expression\.
### 3\.2Formal Definitions

Usage contextA natural language contextuuin which a target expressionxxoccurs\.

Usage instanceA pair\(u,x\)\(u,x\)whereuuis a usage context andxxis a target expression \(a word or multi\-word expression\) occurring inuu\.

Scene representationA structured interpretive representation𝒮​\(u,x\)\\mathcal\{S\}\(u,x\)constructed for usage instance\(u,x\)\(u,x\), consisting of two complementary components: aContextual Scene𝒞\\mathcal\{C\}capturing the broader situational interpretation ofuu, and anExpression Profileℰ\\mathcal\{E\}capturing the scene\-grounded meaning ofxxwithin that context\.

Scene setGiven a set of usage contexts\{ui\}i=1N\\\{u\_\{i\}\\\}\_\{i=1\}^\{N\}containingxx, the corresponding set of scene representations\{𝒮​\(ui,x\)\}i=1N\\\{\\mathcal\{S\}\(u\_\{i\},x\)\\\}\_\{i=1\}^\{N\}constitutes thescene setofxxwith respect to that corpus\. Type\-level properties ofxxcan then be studied through the patterns observed across this collection\.

### 3\.3Contextual Scene

The Contextual Scene𝒞\\mathcal\{C\}represents the broader situational interpretation of a usage instance, independent of which specific expression is targeted\. It consists of three components:Events\(abstracted actions and interactions, expressed using anonymized participant labels such asPersonX,ObjectX\),Entities\(participants and objects salient to the scene, characterized by their role with frame, properties, and emotional state\), andSetting\(spatial, temporal, and atmospheric background\)\. See Table[1](https://arxiv.org/html/2605.22542#S3.T1)for an example\.

While the empirical evaluation in this paper focuses entirely on the Expression Profile, the Contextual Scene is included as part of the full scene representation for its utility in downstream analyses\.

### 3\.4Expression Profile

The Expression Profileℰ\\mathcal\{E\}captures how the target expressionxxis interpreted within the surrounding scene\. Different words in the same sentence will have different Expression Profiles: whatwhiskeyengages in, what properties it bears, and what emotions it evokes in the kitchen scene differs substantially from whattabledoes in the same context\. It consists of three sub\-components\.

##### Engaged Events\.

Events in whichxxplays a central role within the scene\.

##### Generalizable Properties\.

Semantic properties associated withxxin the scene context\.

##### Evoked Emotions\.

Affective associations evoked byxxin the scene context\.

### 3\.5Worked Example

We describe the full LLM\-based generation procedure in Section[4](https://arxiv.org/html/2605.22542#S4)\. Table[1](https://arxiv.org/html/2605.22542#S3.T1)presents a complete scene abstraction for the sentence“The man sat alone at the kitchen table, drinking whiskey late at night,”withwhiskeyas the target expression, generated bygpt\-4o\-mini\.

### 3\.6Scene Similarity

Scene Abstraction comes with a natural notion of scene similarity: given two usage instancesuiu\_\{i\}anduju\_\{j\}, the similarity between𝒮​\(ui,x\)\\mathcal\{S\}\(u\_\{i\},x\)and𝒮​\(uj,x\)\\mathcal\{S\}\(u\_\{j\},x\)provides a structured, interpretable basis for comparing howxxis used across contexts\. Each scene representation can be mapped to a dense vector by embedding its components using a pretrained sentence encoder\. This can be done either by computing a single overall embedding across all components, or by embedding individual components separately, allowing selective focus on specific dimensions of the scene such as events, properties, or evoked emotions\. Given a scene set\{𝒮​\(ui,x\)\}i=1N\\\{\\mathcal\{S\}\(u\_\{i\},x\)\\\}\_\{i=1\}^\{N\}for a wordxx, variation at scene level across contexts can be inspected, making the approach applicable to tasks such as the study of polysemy, contextual variation, sociolinguistic meaning differences, and lexical semantic change\.

## 4Implementation

### 4\.1Overview

Given a usage instance\(u,x\)\(u,x\), our pipeline produces a structured scene representation𝒮​\(u,x\)\\mathcal\{S\}\(u,x\)in two stages: \(1\)scene generation, in which an LLM produces a structured abstraction from the usage context via few\-shot prompting; and \(2\)scene embedding, in which selected components of the structured output are converted into dense vector representations for downstream comparison\. Figure[1](https://arxiv.org/html/2605.22542#S0.F1)illustrates the scene abstraction process111We usegpt\-4o\-minifor all scene generation experiments \(temperature = 0\.2, max\_tokens = 512, top\_p = 1, frequency\_penalty = 0, presence\_penalty = 0\)\.

### 4\.2Scene Generation via Structured Prompting

##### Prompt Design\.

We prompt the model with a system instruction that specifies the desired output format, followed bykkfew\-shot examples that show the intended level of abstraction, and the input sentence with the target expression highlighted\.

Figure[2](https://arxiv.org/html/2605.22542#Ax1.F2)shows the full prompt instruction used for scene abstraction\.

The instruction is governed by four abstraction principles\. These principles are not independent constraints but work in a mutually complementary fashion: the first two govern structural consistency across scenes, while the latter two ensure that each representation retains sufficient information for meaningful semantic interpretation\.

Generalization\.Proper names and specific referring expressions are replaced with role\-based labels \(PersonX,PersonY,ObjectX,AnimalX,PlaceX\), enabling structural similarities to be captured even when the same scene pattern is realized across different surface forms\. This generalization is a prerequisite for analyzing recurring scene structures beyond surface\-level variation\.

Detail Omission\.Narrative details not relevant to the situational interpretation are omitted\. For example,“pulled himself off the wooden floor”is abstracted toPersonX pulls himself off the floor, dropping incidental details such as the material of the floor\.

Interpretability\.All output fields are expressed in natural language that humans can easily read and analyze\. For example, entity properties are written as phrases such as“resilient”or“hopeful yet struggling”rather than numeric identifiers or ontology codes\.

Context Sensitivity\.The Expression Profile is described in terms specific to the given usage context rather than the target expression in general\. For example,stabin“Your deed was another stab, piercing my very heart”yields properties such as“symbolizes emotional pain or betrayal”, whilestabin“my stab at this version of normality”yields“signifies a bold attempt at change”\. Generic descriptions applicable to any context are explicitly discouraged\.

##### Few\-Shot Examples\.

We usek=3k=3few\-shot examples covering keywords with different situational configurations\. Examples were manually constructed to demonstrate the intended level of abstraction\. Table[2](https://arxiv.org/html/2605.22542#S4.T2)presents one such example\.

##### Output Format

The model is instructed to produce a structured output in JSON format covering the components described in Section[3](https://arxiv.org/html/2605.22542#S3)\. Table[1](https://arxiv.org/html/2605.22542#S3.T1)shows a representative example of the output format\.

### 4\.3Scene Embedding

To support vector\-space operations over scene representations, we convert structured scene outputs into dense vector representations via a two\-step procedure\. Because scene representations are structured into distinct components, any subset—whether from the Contextual Scene or the Expression Profile—can be independently encoded, allowing selective focus on the dimensions of meaning most relevant to a given analysis task\. This modularity may offer an interpretability advantage: different components can be compared, aggregated, or weighted independently without retraining\.

In the experiments reported here, we focus on the Expression Profile, which captures the scene\-grounded meaning of the target expression directly and is therefore most amenable to instance\-level semantic comparison\.

Few\-shot example inputFew\-shot example output“Sometimes she would just stay by the window, feeding thecrowswhile he was doing some paperwork, just like old times\.”Keyword:crow

Contextual Scene
\* Events:
PersonX stays near a window; PersonX feeds AnimalGroupX; PersonY does paperwork nearby; PersonX and PersonY share a quiet routine\* Entities:

⋅\\cdotPersonX \(she\): Agent \(Feeding\), Experiencer \(Remembering\_past\); Property: Reflective; Emotion: Calm, Nostalgia
⋅\\cdotPersonY \(he\): Co\-participant \(Routine\_activity\), Worker \(Office\_work\)
⋅\\cdotAnimalGroupX \(the crows\): Recipients \(Feeding\), Symbols \(Memory\_triggering\); Property: Accustomed to being fed\* Setting: ⋅\\cdotPlace: by a window \(indoors\); Time: reflective moment;

⋅\\cdotAtmosphere: calm and nostalgicExpression Profile\(crow = AnimalGroupX\)

⋅\\cdotEngaged events:PersonX feeds them; they receive food as part of a habitual routine
⋅\\cdotGeneralizable properties:commonly present near humans; respond to routine interactions; may evoke memories through repeated presence; passive figures in quiet domestic routines
⋅\\cdotEvoked emotions:Nostalgia \(tied to memories of past shared routines\); Serenity \(presence contributes to a peaceful atmosphere\)Table 2:In\-context learning example forcrow\. The output illustrates how the four abstraction principles operate in practice: role\-based labels \(AnimalGroupX,PersonX\) realize generalization; atmosphere and evoked emotions are inferred rather than stated; and the Expression Profile captures situated rather than dictionary meaning\.##### Textual Serialization\.

Selected scene components are serialized into natural language strings\. For example, theevoked\_emotionsfield\["loneliness", "resignation", "introspection"\]is serialized as“evoked emotions: loneliness, resignation, introspection\.”

##### Embedding\.

Serialized strings are encoded using SentenceBERT \(all\-mpnet\-base\-v2;Reimers and Gurevych \([2019](https://arxiv.org/html/2605.22542#bib.bib25)\)\)\. For the text\-only baseline, raw input text is encoded using the same model to ensure comparability\.

##### Representation Conditions\.

We compare the following representation conditions in Experiment 1 \(Section[5\.1](https://arxiv.org/html/2605.22542#S5.SS1)\):

- •Text: text embedding of the raw input sentence \(baseline\)\.
- •Text \+ Event: original text concatenated with serialized Engaged Events from the expression profile\.
- •Text \+ Property: original text concatenated with serialized Generalizable Properties\.
- •Text \+ Emotion: original text concatenated with serialized Evoked Emotions\.
- •Text \+ Scene: original text concatenated with all three expression profile components\.
- •Scene Only: all three expression profile components without the original text\.

This design allows us to assess both the overall contribution of the expression profile to downstream task performance and the relative discriminative weight of each individual component\.

## 5Experiments

We evaluate Scene Abstraction through two experiments addressing the following research questions:

- •RQ1: Are scene\-level distinctions reliably shared among human observers, and do scene\-based representations capture them better than text embeddings?
- •RQ2: Do LLM\-generated scene abstractions align with human interpretation of scene\-grounded word meaning?

The two experiments reported here focus on the Expression Profileℰ\\mathcal\{E\}\. The Contextual Scene𝒞\\mathcal\{C\}is retained as background to the Expression Profile and as a basis for future analyses\.

### 5\.1Experiment 1: Odd\-Scene\-Out

#### 5\.1\.1Task and Motivation

We evaluate scene\-based representations through anodd\-scene\-outtask: given five usage instances of the same target expression, identify the one whose contextual interpretation is least compatible with the other four\. Scenes have many facets, and different observers may weigh them differently\. However, when a group of usage instances shares a common situational pattern, that shared pattern can be reliably identified even without agreement on a category label\. The odd\-scene\-out design exploits this: rather than asking annotators to categorize scenes, it tests whether they can identify the instance that does not share the situational pattern of the other four\.

#### 5\.1\.2Dataset: COCA\-Scenes

We construct theCOCA\-Scenesdataset to support this experiment\. Twenty\-six keywords were manually selected \(e\.g\.,bathroom,eagle,sword; see Appendix[A COCA\-Scenes Keywords](https://arxiv.org/html/2605.22542#Ax1.SSx1)for the full list\) based on their potential to evoke vivid and varied situational contexts\. For each keyword, four distinct scene types were defined through iterative inspection of corpus examples, aiming for situationally distinct configurations that yield clearly contrasting scene\-level interpretations\. Five sentences per scene type were manually collected from the fiction genre of the Corpus of Contemporary American English\(COCA; Davies,[2008](https://arxiv.org/html/2605.22542#bib.bib26)\)\. COCA is a large, genre\-diverse corpus of American English; its fiction texts were chosen for their rich depictions of events, participants, and atmosphere\.

This yielded a total of 520 sentences \(26 keywords×\\times4 scene types×\\times5 sentences\)\. Each odd\-scene\-out trial is constructed by sampling four sentences from one scene type \(the “base” scenes\) and one sentence from a different scene type for the same keyword \(the “odd” scene\)\. Table[3](https://arxiv.org/html/2605.22542#S5.T3)shows representative stimuli for three keywords\.

Table 3:Representative odd\-scene\-out stimuli\. Each trial presents five sentences sharing a target keyword; participants identify the one depicting a situationally distinct scene \(shown in bold\)\.
#### 5\.1\.3Human Evaluation

Six participants were recruited and divided into two independent groups \(Group 1, Group 2\) of three annotators each\.222This study received an exempt determination from the Institutional Review Board at the University of Texas at Austin \(IRB ID: STUDY00008122\)\.Each participant evaluated 52 trials \(13 keywords×\\times4 trials per keyword\) and selected the sentence depicting the situationally distinct scene\. Every trial was independently judged by all three members within each group, enabling calculation of inter\-annotator agreement \(IAA\)\.

We report Gwet’s AC1\(Gwet,[2008](https://arxiv.org/html/2605.22542#bib.bib27)\)as our primary IAA measure, which estimates chance agreement as a function of how evenly ratings are distributed across categories rather than from marginal frequencies alone, making it robust precisely when one category dominates\. It therefore serves as the primary agreement metric in this study, a decision applied consistently in Experiment 2 as well\.

#### 5\.1\.4Computational Evaluation

In parallel with the human study, we assess whether scene\-based representations provide a discriminative signal for the same task\. We use SentenceBERT \(all\-mpnet\-base\-v2;Reimers and Gurevych \([2019](https://arxiv.org/html/2605.22542#bib.bib25)\)\) to generate sentence embeddings for each candidate\. The candidate with the lowest mean pairwise cosine similarity to the remaining four is selected as the predicted odd\-scene\-out item\. We compare the representation conditions defined in Section[4\.3](https://arxiv.org/html/2605.22542#S4.SS3)\.

#### 5\.1\.5Results

##### Human performance and agreement\.

Human annotators achieved an overall accuracy of82\.37%against the COCA\-Scenes gold standard scene type labels, far exceeding the chance level of 20% \(five\-way choice\)\. Individual accuracy was stable across participants \(range: 73\.08–94\.23%; median: 82\.69%\), indicating that scene\-level distinctions are robust enough to support reliable identification\.

Table[4](https://arxiv.org/html/2605.22542#S5.T4)reports IAA\. Gwet’s AC1 yields a mean score of0\.761\(Substantialagreement, perLandis and Koch \([1977](https://arxiv.org/html/2605.22542#bib.bib28)\)\)\. This is consistent with the view that scene\-level distinctions are reliably identifiable from shared interpretive structure\.

Table 4:Human accuracies and inter\-annotator agreement for Experiment 1 \(Odd\-scene\-out\)\.
##### Computational discriminability\.

Table[5](https://arxiv.org/html/2605.22542#S5.T5)reports the accuracy of model predictions on the odd\-scene\-out task across input conditions\. The text\-only baseline achieves 0\.575\. Integrating the complete set of scene features \(Event, Property, Emotion\) raises accuracy to0\.693, a gain of \+11\.8 percentage points\. The scene\-only condition, which relies solely on abstracted scene features without the original sentence, achieves 0\.627, surpassing the text\-only baseline and demonstrating that abstracted scene features alone carry sufficient semantic weight to identify situational outliers\. Among individual scene components,Propertyproves most discriminative \(0\.661\), suggesting that state\-based and descriptive attributes are particularly central to defining scene boundaries for the concrete nouns in this dataset, though this pattern may not generalize to other word classes\.

Table 5:Automatic evaluation results \(SBERT,all\-mpnet\-base\-v2\) for odd\-scene\-out task\. Event = Engaged Events, Prop = Properties, Emo = Evoked Emotions\.The convergence between human consensus \(AC1 = 0\.761\) and improved computational performance \(Acc = 0\.693\) suggests that the proposed scene schema captures the semantic core of situational distinctions\.

### 5\.2Experiment 2: Human Evaluation of Scene Abstractions

#### 5\.2\.1Setup

##### Task design\.

The second evaluation again focuses on the Expression Profile, this time asking whether humans find the LLM\-generated scene descriptions apt\.

For each keyword\-sentence pair, participants first generated their own open\-ended interpretation through a self\-elicitation step before viewing any model output, responding to the following prompts:

- •Engaged events:“What happened with the \[KEYWORD\] in the situation? What did they do or what occurred to them?”
- •Generalizable properties:“What are the prominent properties of the \[KEYWORD\] in this situation? In your interpretation, what properties stand out as most meaningful or relevant in this context?”
- •Evoked emotions:“Which emotions or wishes does the \[KEYWORD\] evoke in the situation?”

This grounds the evaluation in each participant’s own contextual reading rather than in model\-generated outputs alone\. They then compared two model\-generated representations \(Schema A vs\. Schema B\), selected the one better aligned with their own interpretation, and rated their satisfaction on a 5\-point Likert scale\. If the rating was less than 5, participants identified reasons for dissatisfaction from a predefined checklist:irrelevant information\(not grounded in the situation or unrelated to the keyword\),redundant or verboseinformation,false information\(distorts information from the text\),over\-interpretation\(adds information not supported by the text\),lacks information,hard to understand\(unclear wording or poorly organized\), andother\(offered a free\-text field for elaboration\)\. The self\-elicited prior thus serves as the primary reference point for assessing alignment throughout the analysis\.

##### Stimuli and participants\.

The evaluation set comprised 105 target keywords across 120 usage sentence samples drawn from two sources: DWUG\-EN\(Schlechtweget al\.,[2020](https://arxiv.org/html/2605.22542#bib.bib23)\), a dataset designed for lexical semantic change research that provides diverse usage instances of the same word across varying contexts, and COCA\(Davies,[2008](https://arxiv.org/html/2605.22542#bib.bib26)\)\. The same six annotators from Experiment 1, divided into two groups of three \(Group 1 and 2\), each evaluated 60 keywords across three dimensions \(180 items per participant\), yielding three independent judgments per item\.333The study received an IRB exempt determination \(STUDY00008122\)\.

CategoryRelationDescriptionEngagedEvents\(n=10\)CausesWhat the event causesHinderedByWhat prevents the eventxReasonWhy PersonX actsHasSubEventSteps making up the eventisBeforeWhat usually happens beforeisAfterWhat usually happens afterxIntentWhat PersonX intendsxNeedWhat PersonX needs beforehandxEffectWhat happens to PersonXoEffectWhat happens to othersGen\.Properties\(n=7\)ObjectUseWhat the keyword is used forHasPropertyAttribute or qualityMadeUpOfMaterial or componentsAtLocationWhere typically foundCapableOfWhat the keyword can doDesiresWhat it wants \(if animate\)NotDesiresWhat it avoids \(if animate\)EvokedEmotions\(n=5\)xReactHow PersonX feels afterwardoReactHow others feelxWantWhat PersonX wants nextoWantWhat others wantxAttrHow PersonX is perceived

Table 6:ATOMIC\-2020 relations\(Hwanget al\.,[2021](https://arxiv.org/html/2605.22542#bib.bib18)\)used in ATOMIC Profile, consolidated into three categories matching the evaluation dimensions \(Events:nn=10, Properties:nn=7, Emotions:nn=5\)\.
##### Baseline\.

We compare Schema A \(Scene Profile\) against Schema B \(ATOMIC Profile\) that operationalizes 22 relations in ATOMIC\-2020\(Hwanget al\.,[2021](https://arxiv.org/html/2605.22542#bib.bib18)\), consolidated into three output fields matching the evaluation dimensions \(Table[6](https://arxiv.org/html/2605.22542#S5.T6)\)\. Both schema outputs are generated usinggpt\-4o\-mini\. Schema B provides a principled commonsense baseline: like Schema A, it analyzes a given situation with respect to a target keyword, but derives type\-level generalizations rather than scene\-specific abstractions, making the comparison a direct test of instance\-level versus type\-level semantic representation\.

Table 7:Schema A vs\. Schema B outputs forwhiskey, both generated bygpt\-4o\-mini\. Schema A recovers scene\-conditioned meaning \(solitude, reflection, melancholy\), while Schema B produces type\-general entries \(distilled grains, water, yeast\) applicable in any context\. In Evoked Emotions, Schema A commits to a single scene\-specific label, while Schema B enumerates behavioral\-intentional responses without naming the core affective state\.

#### 5\.2\.2Results

##### Overall preference\.

Scene Profiles were preferred in86\.4%of 1,026 valid evaluations \(360 Engaged Events, 360 Generalizable Properties, 306 Evoked Emotions\), substantially exceeding chance \(p<\.001p<\.001, binomial test for each dimension\)\. When annotators selected Scene Profiles, they did so with high satisfaction \(mean: 4\.78 / 4\.64 / 4\.71 across dimensions, SD≤\\leq0\.78\), significantly higher than when ATOMIC Profiles were selected \(mean: 3\.97 / 4\.33 / 4\.46, SD≤\\leq0\.93; Mann–WhitneyUU,p<\.01p<\.01for all dimensions\)\. Table[8](https://arxiv.org/html/2605.22542#S5.T8)summarizes results by dimension\.

Table 8:Human preference rates for Scene Profiles over ATOMIC Profiles and mean satisfaction ratings \(5\.0 scale\) when each profile type was selected\.
##### Inter\-annotator agreement\.

Gwet’s AC1 yielded substantially higher agreement \(Table[9](https://arxiv.org/html/2605.22542#S5.T9)\)\. Agreement was consistently higher for Group 1 \(AC1: 0\.80–0\.90, Substantial to Almost Perfect\) than Group 2 \(AC1: 0\.32–0\.70, Fair to Substantial\), reflecting annotator\-level variation in evaluation style\.

Table 9:IAA results for Experiment 2\. Full Agr\. = Full Agreement ratio\. Interpretation of Gwet’s AC1 followsLandis and Koch \([1977](https://arxiv.org/html/2605.22542#bib.bib28)\)\.
##### Dimension\-specific findings\.

Scene Profile was preferred over ATOMIC Profiles across all three dimensions, with preference rates of 91\.4% for Engaged Events, 89\.2% for Generalizable Properties, and 77\.1% for Evoked Emotions\. Each dimension is examined below\.

Engaged events\.Scene Profile’s strength lies in recovering event participation that goesbeyond the explicitly stated event structureof the scene\. It correctly infers thatstabin “my stab at this version of normality” denotes an attempt rather than a physical act; thatratin “at the sight of the rat he stiffened and shrank, trembling and shuddered” is identified as a stimulus in a perceptual event \(causing PersonX to react with fear\) even though no rat action is described; and thatravenin sports score data encodes a competing team \(all three annotators, satisfaction: 5 in each case\)\. Of the 31 ATOMIC\-preferred cases \(8\.6%\), 25 \(81%\) were due to insufficient elaboration: sometimes annotators found Scene Profile’s output correct but too sparse\. For example, forpinin“Cousin James, carrying a baseball bat, pinned Jay to the floor”, Scene Profile outputs“PersonX pins PersonY down”, correctly identifying the core action but with minimal elaboration\. ATOMIC Profile outputs“Cousin James uses the baseball bat to pin Jay; Jay is held down”, which one annotator found more complete and descriptive\. In no case did more than one of annotator prefer ATOMIC Profile for Engaged Events\.

Generalizable properties\.Scene Profile captures properties activated by the scene rather than given by encyclopedic defaults\. Forquiltin“Miss Silence had fully worked herself up to the determination of going to the quilting”, Scene Profile outputs“represents a communal activity involving creativity; associated with social gatherings”, while ATOMIC Profile outputs“used for: warmth, decoration, or comfort; made of fabric, often colorful”\. Forratin“at the sight of the rat he stiffened and shrank, trembling and shuddered”, Scene Profile outputs“can evoke strong fear responses in individuals”, while ATOMIC Profile outputs“made of: fur, bones, teeth”\.

Table 10:Cases where ATOMIC Profile was preferred over Scene Profile, with breakdown of annotators’ reasons \(percentages among ATOMIC\-preferred cases per dimension\)\. Each annotator could select multiple reasons, so percentages within a row may sum to more than 100%\.Irrelevant: not grounded in the situation or unrelated to the keyword;Over\-interp\.: adds information not supported by the text;False info\.: distorts information from the text;Verbose: redundant or unnecessarily verbose;Hard to understand: unclear wording or poorly organized;N/A: Scene Profile produced no evoked emotion output for this usage instance\. This category applies only to Evoked Emotions; grayed cells indicate it is not applicable to the other two dimensions\.Evoked emotions\.Scene Profile was preferred less often for Evoked Emotions \(77\.1%\) than for the other two dimensions, which may reflect a structural difference between the two profiles\. Scene Profile is designed to capture the emotion evoked in the reader by the target expression, committing to one interpretive stance\. ATOMIC Profile, by contrast, enumerates the reactions and desires of multiple participants within the scene \(xReact,oReact,xWant,oWant\), which may align more readily with annotators who foregrounded different participants\. Forbeerin“he had been training for months and habitually refusing my offers to bring over beer on the weekends”, annotators described the evoked emotion as “wish to socialize”, “beer is not wanted here”, and “socialization”, each reflecting a different participant’s perspective\. Scene Profile output“Disappointment”, aligning with one reading but not others, while ATOMIC Profile output“wants to: avoid distractions from training; others want to: socialize and enjoy drinks together”, accommodating multiple perspectives\. All three annotators preferred ATOMIC Profile in this case\.

##### Failure case analysis\.

Table[10](https://arxiv.org/html/2605.22542#S5.T10)summarizes failure modes across dimensions\. Scene Profile’s failures largely fall into two patterns\. Inlacks\-infocases, Scene Profile correctly identified the scene role but did not elaborate it sufficiently: forwinein“we sit outside at a table with a checked cloth and drink wine in the sun”\(Engaged Events\), Scene Profile output only“PersonX and PersonY drink it”, which annotators found too sparse\. Inover\-interpretationcases, Scene Profile inferred a property not warranted by the scene: forlavenderin“it had like a purple satin ruffle skirt, maybe lavender”\(Generalizable Properties\), Scene Profile output“may evoke a sense of calm or tranquility”, while an annotator’s own response was simply“it’s the color of the skirt”\. For Evoked Emotions, Scene Profile sometimes producedNonewhen it identified no specific emotions evoked by the target expression\. Annotators who disagreed with this judgment found the output unsatisfying \(24% of ATOMIC\-preferred Emotion cases; see N/A column in Table[10](https://arxiv.org/html/2605.22542#S5.T10)\)\. The over\-interpretation rate was consistently higher when annotators reported “nothing specific” in their self\-elicited response \(Property: 33\.3% vs\. 10\.1%; Emotion: 20\.7% vs\. 7\.9%\), suggesting that Scene Profile over\-interprets most when the scene itself provides insufficient grounding for a specific inference\.

##### Qualitative feedback\.

At the end of the experiment, annotators provided free\-text assessments of each profile’s overall strengths and weaknesses \(see Table[12](https://arxiv.org/html/2605.22542#Ax1.T12)in Appendix C for full responses\)\. Scene Profile was consistently characterized as more concise and contextually accurate, though some annotators noted a tendency to over\-interpret or to summarize the scene rather than focus on the target word\. ATOMIC Profile was described as more thorough and complete, but frequently criticized for being verbose, redundant, and including information irrelevant to the specific usage context\.

## 6Discussion

### 6\.1What Scene Representations Capture

The results of Experiment 1 provide converging evidence from two independent evaluation tracks that scene\-level distinctions constitute a meaningful unit among human participants\. They identified outlier situations of target word instances with 82\.37% accuracy, far above the 20% chance level, and with substantial inter\-annotator consensus \(Gwet’s AC1 = 0\.761\)\. This suggests that scene\-level distinctions reflect shared perceptual\-interpretive structure rather than task\-specific conventions\.

The computational results mirror this pattern\. Integrating scene features raises odd\-scene\-out accuracy from 0\.575 to 0\.693 \(\+11\.8 percentage points\), and the scene\-only condition alone outperforms the text baseline \(0\.627 vs\. 0\.575\)\. Scene abstractions derived entirely from LLM\-generated structured representations, with no access to the original sentence, carry sufficient semantic weight to distinguish situational outliers\. This suggests that scene features may reflect a distinct layer of situational interpretation rather than a restatement of surface lexical content\.

Among individual scene components in the odd\-scene\-out task,Generalizable Propertiesemerges as the most discriminative feature \(0\.661\), outperforming both Engaged Events \(0\.609\) and Evoked Emotions \(0\.562\)\. This pattern may reflect the nature of the keywords chosen in COCA\-Scenes, which are predominantly concrete nouns whose situational configurations are particularly well distinguished by property\-level attributes\.

### 6\.2Scene\-Grounding as a Mechanism of Alignment

As reported in Section[5\.2](https://arxiv.org/html/2605.22542#S5.SS2), Scene Profiles were preferred in 86\.4% of evaluations across all three dimensions\. This advantage stems not from general output quality but from a representational difference: Scene Profile conditions its abstraction on the particular scene instance\. ATOMIC Profile, by contrast, is organized around relational schemas that describe what typically holds for a keyword in general; as shown in Table[7](https://arxiv.org/html/2605.22542#S5.T7), this can result in outputs that apply broadly across contexts rather than capturing what is specific to the scene at hand\.

Lacks\-infowas the most common reason for preferring ATOMIC Profile in Engaged Events \(77% of ATOMIC\-preferred cases\), whileover\-interpretationwas more prominent in Generalizable Properties \(28%\) and Evoked Emotions \(36%\)\. Over\-interpretation rates were consistently higher when annotators wrote “nothing specific” in their self\-elicitation response \(Property: 33\.3% vs\. 10\.1%; Emotion: 20\.7% vs\. 7\.9%\), suggesting that Scene Profile is prone to overreaching or hallucinating when the usage instance provides little scene\-relevant information to draw from\.

## 7Conclusion

We introduced Scene Abstraction, a computational approach that constructs structured scene representations𝒮​\(u,x\)\\mathcal\{S\}\(u,x\)for word usage instances\. Each scene comprises a Contextual Scene𝒞\\mathcal\{C\}capturing the broader situational interpretation and an Expression Profileℰ\\mathcal\{E\}capturing the scene\-grounded meaning of a target expression, operationalized via few\-shot prompting of an LLM\.

Two experiments provide complementary support for the framework\. Experiment 1 shows that scene\-level distinctions are reliably identifiable across human observers \(82\.37% accuracy, AC1 = 0\.761\) and that scene\-based representations outperform raw\-text embeddings on this task \(\+11\.8 pp\)\. Experiment 2 shows that LLM\-generated scene profiles align more closely with human scene\-grounded word meaning than ATOMIC\-based commonsense profiles \(86\.4% preference\)\.

Together, these results suggest that scene representations reflect aspects of how humans interpret words in situated contexts\. The three evaluation dimensions \(what event a word engages in, what properties it bears, and what emotion it evokes\) show different patterns of human agreement, pointing to the complexity of scene\-grounded interpretation\. Scene Abstraction offers one way to make aspects of this interpretation explicit and computationally accessible, complementing existing distributional and knowledge\-based approaches\.

Several directions emerge naturally from this work\. Extending the framework to metaphorical and non\-literal language is a natural next step, as metaphor often arises from structural and property\-level similarities across situations, dimensions that scene representations capture explicitly\. Aggregating scene representations across usage instances to study a type\-level semantics of a word is another direction worth exploring\.

### Limitations

The framework has several limitations worth noting\. Scene abstractions are generated by a prompted LLM and may inherit systematic biases or hallucinated inferences from the underlying model\. The affective and interpretive dimensions of scenes \(inferred properties and emotions, atmosphere, evoked emotions\) are understood as subjective inferences from situational context rather than fixed, objective semantic properties, and the current evaluation design does not fully control for this subjectivity\. We make no claim that the scene representations produced by this framework correspond directly to cognitive representations in the psychological sense; the framework is a representational and analytical tool, not a model of mental processing\. Additionally, the framework has been validated primarily on English\-language fiction; generalizability across languages, genres, and cultural contexts remains to be assessed\.

## References

- C\. F\. Baker, C\. J\. Fillmore, and J\. B\. Lowe \(1998\)The Berkeley FrameNet project\.InCOLING 1998 Volume 1: The 17th International Conference on Computational Linguistics,External Links:[Link](https://aclanthology.org/C98-1013/)Cited by:[§1](https://arxiv.org/html/2605.22542#S1.p2.1),[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px1.p1.1)\.
- A\. Bosselut, H\. Rashkin, M\. Sap, C\. Malaviya, A\. Celikyilmaz, and Y\. Choi \(2019\)COMET: commonsense transformers for automatic knowledge graph construction\.InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics,A\. Korhonen, D\. Traum, and L\. Màrquez \(Eds\.\),Florence, Italy,pp\. 4762–4779\.External Links:[Link](https://aclanthology.org/P19-1470/),[Document](https://dx.doi.org/10.18653/v1/P19-1470)Cited by:[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px4.p1.1)\.
- T\. Brown, B\. Mann, N\. Ryder, M\. Subbiah, J\. D\. Kaplan, P\. Dhariwal, A\. Neelakantan, P\. Shyam, G\. Sastry, A\. Askell, S\. Agarwal, A\. Herbert\-Voss, G\. Krueger, T\. Henighan, R\. Child, A\. Ramesh, D\. Ziegler, J\. Wu, C\. Winter, C\. Hesse, M\. Chen, E\. Sigler, M\. Litwin, S\. Gray, B\. Chess, J\. Clark, C\. Berner, S\. McCandlish, A\. Radford, I\. Sutskever, and D\. Amodei \(2020\)Language models are few\-shot learners\.InAdvances in Neural Information Processing Systems,H\. Larochelle, M\. Ranzato, R\. Hadsell, M\.F\. Balcan, and H\. Lin \(Eds\.\),Vol\.33,pp\. 1877–1901\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2020/file/1457c0d6bfcb4967418bfb8ac142f64a-Paper.pdf)Cited by:[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px5.p1.1)\.
- N\. Chambers and D\. Jurafsky \(2008\)Unsupervised learning of narrative event chains\.InProceedings of ACL\-08: HLT,J\. D\. Moore, S\. Teufel, J\. Allan, and S\. Furui \(Eds\.\),Columbus, Ohio,pp\. 789–797\.External Links:[Link](https://aclanthology.org/P08-1090/)Cited by:[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px2.p1.1)\.
- M\. Davies \(2008\)The Corpus of Contemporary American English \(COCA\)\.Available online at https://www\.english\-corpora\.org/coca/\.Cited by:[§5\.1\.2](https://arxiv.org/html/2605.22542#S5.SS1.SSS2.p1.1),[§5\.2\.1](https://arxiv.org/html/2605.22542#S5.SS2.SSS1.Px2.p1.1)\.
- J\. Devlin, M\. Chang, K\. Lee, and K\. Toutanova \(2019\)BERT: pre\-training of deep bidirectional transformers for language understanding\.InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 \(Long and Short Papers\),J\. Burstein, C\. Doran, and T\. Solorio \(Eds\.\),Minneapolis, Minnesota,pp\. 4171–4186\.External Links:[Link](https://aclanthology.org/N19-1423/),[Document](https://dx.doi.org/10.18653/v1/N19-1423)Cited by:[§1](https://arxiv.org/html/2605.22542#S1.p2.1),[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px3.p1.1)\.
- C\. J\. Fillmore \(1976\)Frame semantics and the nature of language\.Annals of the New York Academy of Sciences280\(1\),pp\. 20–32\.Cited by:[§1](https://arxiv.org/html/2605.22542#S1.p2.1),[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px1.p1.1)\.
- C\. J\. Fillmore \(1982\)Frame semantics\.InLinguistics in the Morning Calm,pp\. 111–137\.Cited by:[§1](https://arxiv.org/html/2605.22542#S1.p2.1),[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px1.p1.1)\.
- F\. Gilardi, M\. Alizadeh, and M\. Kubli \(2023\)ChatGPT outperforms crowd workers for text\-annotation tasks\.Proceedings of the National Academy of Sciences120\(30\),pp\. e2305016120\.Cited by:[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px5.p1.1)\.
- A\. E\. Goldberg \(1995\)Constructions: a construction grammar approach to argument structure\.University of Chicago Press,Chicago\.Cited by:[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px1.p1.1)\.
- Y\. Gu, B\. Dalvi, and P\. Clark \(2022\)DREAM: improving situational QA by first elaborating the situation\.InProceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,M\. Carpuat, M\. de Marneffe, and I\. V\. Meza Ruiz \(Eds\.\),Seattle, United States,pp\. 1115–1127\.External Links:[Link](https://aclanthology.org/2022.naacl-main.82/),[Document](https://dx.doi.org/10.18653/v1/2022.naacl-main.82)Cited by:[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px5.p1.1)\.
- K\. L\. Gwet \(2008\)Computing inter\-rater reliability and its variance in the presence of high agreement\.British Journal of Mathematical and Statistical Psychology61\(1\),pp\. 29–48\.Cited by:[§5\.1\.3](https://arxiv.org/html/2605.22542#S5.SS1.SSS3.p2.1)\.
- Z\. S\. Harris \(1954\)Distributional structure\.Word10\(2\-3\),pp\. 146–162\.Cited by:[§1](https://arxiv.org/html/2605.22542#S1.p2.1),[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px3.p1.1)\.
- A\. Hoyle, R\. Sarkar, P\. Goel, and P\. Resnik \(2023\)Natural language decompositions of implicit content enable better text representations\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,H\. Bouamor, J\. Pino, and K\. Bali \(Eds\.\),Singapore,pp\. 13188–13214\.External Links:[Link](https://aclanthology.org/2023.emnlp-main.815/),[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.815)Cited by:[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px3.p1.1)\.
- J\. D\. Hwang, C\. Bhagavatula, R\. Le Bras, J\. Da, K\. Sakaguchi, A\. Bosselut, and Y\. Choi \(2021\)\(Comet\-\) atomic 2020: on symbolic and neural commonsense knowledge graphs\.Proceedings of the AAAI Conference on Artificial Intelligence35\(7\),pp\. 6384–6392\.External Links:[Link](https://ojs.aaai.org/index.php/AAAI/article/view/16792),[Document](https://dx.doi.org/10.1609/aaai.v35i7.16792)Cited by:[§1](https://arxiv.org/html/2605.22542#S1.p4.1),[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px4.p1.1),[§5\.2\.1](https://arxiv.org/html/2605.22542#S5.SS2.SSS1.Px3.p1.1),[Table 6](https://arxiv.org/html/2605.22542#S5.T6)\.
- T\. K\. Landauer and S\. T\. Dumais \(1997\)A solution to plato’s problem: the latent semantic analysis theory of acquisition, induction, and representation of knowledge\.\.Psychological review104\(2\),pp\. 211\.Cited by:[§1](https://arxiv.org/html/2605.22542#S1.p2.1)\.
- J\. R\. Landis and G\. G\. Koch \(1977\)The measurement of observer agreement for categorical data\.Biometrics,pp\. 159–174\.Cited by:[§5\.1\.5](https://arxiv.org/html/2605.22542#S5.SS1.SSS5.Px1.p2.1),[Table 9](https://arxiv.org/html/2605.22542#S5.T9)\.
- I\. Lee and D\. Goldwasser \(2018\)FEEL: featured event embedding learning\.Proceedings of the AAAI Conference on Artificial Intelligence32\(1\)\.External Links:[Link](https://ojs.aaai.org/index.php/AAAI/article/view/11936),[Document](https://dx.doi.org/10.1609/aaai.v32i1.11936)Cited by:[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px3.p1.1)\.
- N\. Mostafazadeh, A\. Kalyanpur, L\. Moon, D\. Buchanan, L\. Berkowitz, O\. Biran, and J\. Chu\-Carroll \(2020\)GLUCOSE: GeneraLized and COntextualized story explanations\.InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),B\. Webber, T\. Cohn, Y\. He, and Y\. Liu \(Eds\.\),Online,pp\. 4569–4586\.External Links:[Link](https://aclanthology.org/2020.emnlp-main.370/),[Document](https://dx.doi.org/10.18653/v1/2020.emnlp-main.370)Cited by:[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px4.p1.1)\.
- M\. E\. Peters, M\. Neumann, M\. Iyyer, M\. Gardner, C\. Clark, K\. Lee, and L\. Zettlemoyer \(2018\)Deep contextualized word representations\.InProceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 \(Long Papers\),M\. Walker, H\. Ji, and A\. Stent \(Eds\.\),New Orleans, Louisiana,pp\. 2227–2237\.External Links:[Link](https://aclanthology.org/N18-1202/),[Document](https://dx.doi.org/10.18653/v1/N18-1202)Cited by:[§1](https://arxiv.org/html/2605.22542#S1.p2.1),[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px3.p1.1)\.
- K\. Pichotta and R\. J\. Mooney \(2016\)Using sentence\-level LSTM language models for script inference\.InProceedings of the 54th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),K\. Erk and N\. A\. Smith \(Eds\.\),Berlin, Germany,pp\. 279–289\.External Links:[Link](https://aclanthology.org/P16-1027/),[Document](https://dx.doi.org/10.18653/v1/P16-1027)Cited by:[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px2.p1.1)\.
- N\. Reimers and I\. Gurevych \(2019\)Sentence\-BERT: sentence embeddings using Siamese BERT\-networks\.InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing \(EMNLP\-IJCNLP\),K\. Inui, J\. Jiang, V\. Ng, and X\. Wan \(Eds\.\),Hong Kong, China,pp\. 3982–3992\.External Links:[Link](https://aclanthology.org/D19-1410/),[Document](https://dx.doi.org/10.18653/v1/D19-1410)Cited by:[§4\.3](https://arxiv.org/html/2605.22542#S4.SS3.SSS0.Px2.p1.1),[§5\.1\.4](https://arxiv.org/html/2605.22542#S5.SS1.SSS4.p1.1)\.
- M\. Sap, R\. Le Bras, E\. Allaway, C\. Bhagavatula, N\. Lourie, H\. Rashkin, B\. Roof, N\. A\. Smith, and Y\. Choi \(2019\)ATOMIC: an atlas of machine commonsense for if\-then reasoning\.Proceedings of the AAAI Conference on Artificial Intelligence33\(01\),pp\. 3027–3035\.External Links:[Link](https://ojs.aaai.org/index.php/AAAI/article/view/4160),[Document](https://dx.doi.org/10.1609/aaai.v33i01.33013027)Cited by:[§1](https://arxiv.org/html/2605.22542#S1.p4.1),[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px4.p1.1)\.
- R\.C\. Schank and R\.P\. Abelson \(1977\)Scripts, plans, goals, and understanding: an inquiry into human knowledge structures\.Artificial intelligence series,L\. Erlbaum Associates\.External Links:ISBN 9780470990339,LCCN lc76051963,[Link](https://books.google.com/books?id=YZ99AAAAMAAJ)Cited by:[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px2.p1.1)\.
- D\. Schlechtweg, B\. McGillivray, S\. Hengchen, H\. Dubossarsky, and N\. Tahmasebi \(2020\)SemEval\-2020 task 1: unsupervised lexical semantic change detection\.InProceedings of the Fourteenth Workshop on Semantic Evaluation,A\. Herbelot, X\. Zhu, A\. Palmer, N\. Schneider, J\. May, and E\. Shutova \(Eds\.\),Barcelona \(online\),pp\. 1–23\.External Links:[Link](https://aclanthology.org/2020.semeval-1.1/),[Document](https://dx.doi.org/10.18653/v1/2020.semeval-1.1)Cited by:[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px3.p1.1),[§5\.2\.1](https://arxiv.org/html/2605.22542#S5.SS2.SSS1.Px2.p1.1)\.
- T\. A\. Van Dijk, W\. Kintsch,et al\.\(1983\)Strategies of discourse comprehension\.Academic press New York\.Cited by:[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px2.p1.1)\.
- Z\. Wang, Q\. Xie, Y\. Feng, Z\. Ding, Z\. Yang, and R\. Xia \(2024\)Is chatgpt a good sentiment analyzer? a preliminary study\.External Links:2304\.04339,[Link](https://arxiv.org/abs/2304.04339)Cited by:[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px5.p1.1)\.
- J\. Wei, X\. Wang, D\. Schuurmans, M\. Bosma, b\. ichter, F\. Xia, E\. Chi, Q\. V\. Le, and D\. Zhou \(2022\)Chain\-of\-thought prompting elicits reasoning in large language models\.InAdvances in Neural Information Processing Systems,S\. Koyejo, S\. Mohamed, A\. Agarwal, D\. Belgrave, K\. Cho, and A\. Oh \(Eds\.\),Vol\.35,pp\. 24824–24837\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2022/file/9d5609613524ecf4f15af0f7b31abca4-Paper-Conference.pdf)Cited by:[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px5.p1.1)\.
- P\. West, C\. Bhagavatula, J\. Hessel, J\. D\. Hwang, L\. Jiang, R\. Le Bras, X\. Lu, S\. Welleck, and Y\. Choi \(2022\)Symbolic knowledge distillation: from general language models to commonsense models\.InProceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,M\. Carpuat, M\. de Marneffe, and I\. V\. Meza Ruiz \(Eds\.\),Seattle, United States,pp\. 4602–4625\.External Links:[Link](https://aclanthology.org/2022.naacl-main.341/),[Document](https://dx.doi.org/10.18653/v1/2022.naacl-main.341)Cited by:[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px5.p1.1)\.
- G\. Wiedemann, S\. Remus, A\. Chawla, and C\. Biemann \(2019\)Does BERT make any sense? interpretable word sense disambiguation with contextualized embeddings\.CoRRabs/1909\.10430\.External Links:[Link](http://arxiv.org/abs/1909.10430),1909\.10430Cited by:[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px3.p1.1)\.
- R\. A\. Zwaan, J\. P\. Magliano, and A\. C\. Graesser \(1995\)Dimensions of situation model construction in narrative comprehension\.\.Journal of experimental psychology: Learning, memory, and cognition21\(2\),pp\. 386\.Cited by:[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px2.p1.1)\.
- R\. A\. Zwaan and G\. A\. Radvansky \(1998\)Situation models in language comprehension and memory\.\.Psychological bulletin123\(2\),pp\. 162\.Cited by:[§2](https://arxiv.org/html/2605.22542#S2.SS0.SSS0.Px2.p1.1)\.

## Appendix

### A COCA\-Scenes Keywords

Table[11](https://arxiv.org/html/2605.22542#Ax1.T11)lists the 26 keywords used in the COCA\-Scenes dataset\.

bathroombeerchickencigarcigarettecoffeecroweagleelevatorfireknifemicrowaveovenowlpistolraccoonratravenriflesquirrelswordteaturkeyvulturewaterwine
Table 11:The 26 keywords in the COCA\-Scenes dataset\.
### B Prompt Instruction

Figure[2](https://arxiv.org/html/2605.22542#Ax1.F2)shows the full prompt instruction used for scene abstraction\.

Prompt: Scene Abstraction InstructionsGiven a sentence and a target keyword, extract a structured scene abstraction and format your response as a bullet point list, covering the following fields:1\. Events:\-List only what happened in COMET\-ATOMIC style \(e\.g\., “PersonX opens ObjectY”\)\.\-Do not include properties, emotions, or interpretations\.\-Do not include definitions or static facts\.2\. Entities:\-List each entity with its surface mention \(e\.g\., “PersonX \(she\)”\)\.\-For each entity, include:∘\\circRole\(s\) with associated Frame\(s\)∘\\circProperty: context\-specific traits with brief explanations∘\\circEmotion: inferred emotional states with explanations3\. Setting:\-Describe Place, Time, and Atmosphere if inferable\.\-Use “unspecified” when not identifiable\.4\. Expression Profile:\-Keyword: the target linguistic expression \(with its assigned label\)\-Engaged Events: COMET\-ATOMIC style events involving the keyword\-Generalizable Properties: abstract traits implied by the contextConstraints:•All entity labels must use X/Y/Z suffixes only \(e\.g\., PersonX\)\.•Format must be a bullet\-style list without extra explanation\.

Figure 2:The prompt instruction for the scene abstraction process\.

### C Annotator Qualitative Feedback

Table 12:Annotator free\-text assessments ofScene Profile \(A\)andATOMIC Profile \(B\), collected at the end of Experiment 2\. Annotators 1–3 form Group 1 and Annotators 4–6 form Group 2\.