💬 NLP

What makes an entity salient in discourse?

This paper investigates how utterance-level prominence predictors interact with entity-level factors and genre variations to determine discourse-level salience, finding that while grammatical features correlate with salience, discourse-structural and semantic factors are more robust determinants across 24 English genres.

Original authors: Amir Zeldes, Jessica Lin

Published 2026-04-21

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Amir Zeldes, Jessica Lin

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 at a massive, chaotic party with hundreds of people talking at once. Some people are the life of the party, the ones everyone keeps talking about. Others are just background noise, mentioned once and then forgotten.

This paper is essentially a detective story trying to figure out what makes a person (or object) the "star" of the conversation versus just a cameo appearance. The authors, Amir Zeldes and Jessica Lin, wanted to answer a big question: When we read a story or listen to a speech, how do we decide what is important enough to remember and what can be left out?

Here is the breakdown of their investigation, using some simple analogies.

1. The Problem: How do we measure "Importance"?

In the past, linguists tried to guess importance by looking at grammar rules. They thought, "If someone is the subject of a sentence (like 'The dog barked'), they must be important." Or, "If someone is a human, they are more important than a rock."

But the authors realized this is like judging a movie by only looking at who is standing in the center of the frame. Sometimes the main character is in the background, and sometimes a rock is the most important thing in the scene (think of a magic ring in Lord of the Rings).

To solve this, they invented a new way to measure importance called "Summary-Worthiness."

The Analogy: Imagine you have to summarize a 300-page book onto a single postcard. You can't write everything down. You have to pick the absolute most critical things.
The Method: They took 280 different texts (everything from news articles and court transcripts to travel guides and fiction) and asked humans to write summaries. If an entity (a person, place, or thing) appeared in the summary, it got a "salience" point. If it appeared in 5 different summaries of the same text, it was a Superstar. If it appeared in none, it was Background Noise.

2. The Investigation: What actually makes a Star?

They ran a massive computer analysis to see which features predicted whether an entity would end up in the summary. They looked at three main categories of clues:

A. The "Grammar Clues" (The Surface Level)

Old Theory: "If they are the subject of the sentence, they are important."
The Reality: It helps, but it's not the whole story. Being a subject is like wearing a nice suit; it helps you stand out, but if you are just a minor character in a suit, you still get cut from the summary.
Surprise: Sometimes, being a "possessive" (like John's book) or a "vocative" (calling someone by name, like "Hey, John!") made an entity more likely to be important than just being a subject.

B. The "Character Clues" (Who are they?)

Old Theory: "Humans are always more important than rocks."
The Reality: This is true usually, but not always.
- The Travel Guide Exception: In a travel guide about Paris, the city of Paris is the star. The people mentioned (like a baker or a tourist) are just supporting actors. If you wrote a summary of a travel guide, you'd write about the city, not the baker.
- The Lesson: Importance depends on the genre (the type of text). You can't have a single rule for all texts.

C. The "Structure Clues" (Where do they fit in the story?)

This was the most powerful part of their discovery. They looked at how the text is built, like the architecture of a building.

The "Tree" Analogy: Imagine the text is a tree. The trunk and main branches are the most important parts (the main points). The tiny twigs at the very top are the details.
- They found that if an entity appears on the main trunk of the story, it is much more likely to be important.
- If an entity is buried deep in a "twig" (a side story or an explanation), it's less likely to be remembered.
The "Spread" Analogy: Think of an entity like a rumor.
- If a rumor is only told once in one corner of the room, it dies out.
- If a rumor is told many times (frequency) and all over the room (dispersion), it becomes the main topic.
- The authors found that how widely an entity is spread out across the text is the single strongest predictor of importance.

3. The Big Reveal

The authors built a "Super-Model" (a computer program) to predict importance. Here is what they learned:

It's not just one thing: You can't just look at grammar or just look at whether someone is a human. It's a mix of everything.
Context is King: The type of text matters more than you think. A "human" is a star in a biography but a side character in a travel guide.
The "Spread" Wins: The most reliable sign of importance is persistence. If an entity keeps showing up throughout the whole document, it's a star. If it shows up once and vanishes, it's likely forgettable.
Structure Matters: Where an entity sits in the "hierarchy" of the story (is it the main point or a side note?) is a huge clue.

The Takeaway

This paper teaches us that importance isn't a fixed property of a word; it's a relationship between the word, the story, and the reader's goal.

If you are reading a recipe, the "ingredients" are the stars.
If you are reading a novel, the "characters" are the stars.
If you are reading a travel guide, the "places" are the stars.

The authors successfully mapped out the rules of this game, showing us that to understand what makes something important in a conversation, we have to look at the whole picture, not just the individual sentences. They proved that our brains (and computers) are very good at spotting the "main trunk" of the story and ignoring the "twigs," but only if we give them the right context.

1. Problem Statement

The paper addresses the challenge of defining and operationalizing discourse-level salience—the degree to which an entity (e.g., a person, object, or abstract concept) remains prominent and memorable throughout a text, distinct from utterance-level salience (local prominence within a single sentence).

While previous research has extensively studied local prominence (e.g., via Centering Theory, which focuses on grammatical subjects and pronouns), there is a lack of understanding regarding:

How local predictors translate to global discourse salience.
How salience varies across diverse genres (spoken vs. written, formal vs. informal).
How to measure salience consistently without relying on circular linguistic definitions or subjective human judgments that vary by annotator.

2. Methodology

Data Source

The authors utilize the Georgetown University Multilayer (GUM) corpus and the GENTLE (GENre Tests for Linguistic Evaluation) partition.

Scope: ~268,000 tokens across 281 documents.
Genres: 24 distinct genres including academic writing, biographies, vlogs, conversations, news, fiction, travel guides, legal documents, and mathematical proofs.
Annotations: The data is richly annotated with Universal Dependencies (syntax), coreference resolution, and Enhanced Rhetorical Structure Theory (eRST) discourse parsing.

Operationalization of Salience

To avoid subjective bias, the authors define salience based on summary-worthiness:

Metric: An entity's salience score is the number of times it appears in a set of multiple summaries (ranging from 0 to 5).
Data Collection: Each document has 5 summaries (a mix of human-written and LLM-generated, manually corrected).
Validation: Inter-annotator agreement on whether an entity appears in a summary was extremely high ( $\kappa = 0.9782$ ), validating the metric.
Scale: Salience is treated as a graded, discrete variable (0–5), following a U-shaped distribution.

Analytical Approach

The study employs a two-stage modeling approach:

Single-Layer Analyses: Univariate and multivariate linear mixed-effects models (using beta-binomial regression) to test the significance of individual predictors.
Multilayer Predictive Modeling: A Random Forest ensemble (Extra Trees algorithm) to capture non-linear interactions and evaluate generalization across genres.
Feature Categories:
- Morphosyntactic: Grammatical function (subjecthood), part-of-speech, definiteness, number, linear position.
- Entity-Based: Entity type (animacy, e.g., person, place, abstract), cluster size (mention frequency), and dispersion (Kullback-Leibler divergence across the document).
- Discourse Structure: eRST nesting depth, discourse relation types (e.g., elaboration, attribution), and explicit discourse markers.
- Referential Structure: Centering Theory metrics (Cf rankings, Cb transitions, continuation vs. shift).

3. Key Contributions

Graded Operationalization: Introduction of a robust, extra-textual metric for discourse salience based on multi-summary alignment, applicable to any text type (spoken or written).
Comprehensive Corpus Analysis: The first large-scale analysis of discourse-level salience covering 24 genres with rich annotations for syntax, coreference, and discourse structure.
Interaction of Factors: Demonstration that while utterance-level predictors (like subjecthood) correlate with discourse salience, they are modulated by entity-level factors (frequency, dispersion) and genre-specific communicative intents.
Predictive Modeling: Development of interpretable models that identify robust indicators of salience and demonstrate generalization to out-of-domain genres (GENTLE).

4. Key Results

A. Morphosyntax vs. Discourse Structure

Subjecthood is not the sole driver: While grammatical subjects are more salient than other functions, they are not the most salient category. Possessives, vocatives, and indirect objects often show higher mean salience scores. This is largely because these roles are disproportionately filled by human participants.
Discourse Structure is Robust: Features derived from discourse structure (eRST depth, discourse relations) and referential structure (Centering transitions) proved more robust predictors than simple morphosyntactic features.
- Nesting Depth: Entities appearing in central (shallow depth) discourse units are significantly more salient.
- Discourse Relations: Entities in "Attribution," "Organization" (headings), and "Topic" relations are highly salient. Conversely, "Elaboration" and "Explanation" relations tend to contain less salient entities.

B. Entity-Level Factors

Frequency and Dispersion: The number of mentions (cluster size) and the dispersion of mentions across the document (Kullback-Leibler divergence) are the strongest predictors of salience. An entity mentioned frequently but clustered in one section is less salient than one dispersed throughout the text.
Animacy and Genre Interaction:
- General Trend: Humans (Person entities) are generally the most salient.
- Genre Exception: In Travel Guides, Places are more salient than people. In How-to Guides, generic "you" is often non-salient, whereas in Conversations, "I/You" are highly salient.
- Conclusion: No single entity type is universally salient; the "salience profile" depends heavily on the genre's communicative intent.

C. Predictive Modeling Performance

Variable Importance: In the Random Forest model, Cluster Divergence (dispersion), Cluster Size, and Minimum Discourse Depth were the top predictors.
Genre Importance: Genre was a significant predictor (ranking 5th), but the model generalized well to out-of-domain data (GENTLE) with only a ~3% drop in accuracy. This suggests that top-level structural features (dispersion, depth) are universal, while genre acts as a modulator for specific entity types.
Performance: The model achieved 83.6% accuracy on the in-domain test set and 80.9% on the out-of-domain set (baseline ~60-63%).

5. Significance and Implications

Theoretical Validation: The study provides empirical validation for theories like Centering Theory and Rhetorical Structure Theory (RST). It confirms that constructs developed for local coherence (e.g., Cb transitions, nuclearity) are predictive of global discourse importance, bridging the gap between sentence-level and document-level analysis.
Beyond "Subjecthood": The paper challenges the traditional linguistic assumption that grammatical subjecthood is the primary marker of prominence. It argues that semantic and pragmatic factors (animacy, discourse function, and distribution) are more critical for determining what humans remember and summarize.
Application to NLP: The findings offer a blueprint for improving Automatic Summarization and Coreference Resolution systems. By incorporating discourse depth and entity dispersion, models can better identify "summary-worthy" content across diverse genres, moving beyond simple frequency counts or syntactic heuristics.
Methodological Rigor: The use of multiple summaries to create a graded salience score offers a replicable, less subjective alternative to direct human annotation of importance, setting a new standard for future salience research.

In conclusion, the paper establishes that discourse-level salience is a multifactorial phenomenon driven by a complex interplay of how often an entity appears, where it appears in the discourse structure, what type of entity it is, and what genre the text belongs to. No single linguistic feature (like subjecthood) is sufficient; rather, it is the convergence of structural and semantic cues that determines prominence.