Structural Prominence Hierarchies for Discourse Salience

88 concepts63 papers1 proposal

When humans read or listen to a sentence, they do not treat all entities equally. Consider the sequence: "The chef prepared the meal. He..." Instantly, our minds map "He" to the chef, not the meal. Why? It isn't because a meal is semantically incapable of performing actions (though semantics certainly helps); it is because the speaker structured the sentence to place the chef at the center of our attention.

In natural language processing, tracking this center of attention is known as modeling discourse salience. While early computational frameworks attempted to track salience using exhaustive world knowledge or complex semantic frames, a long line of research demonstrates that language provides a highly efficient shortcut: structural prominence hierarchies. By ranking entities based on their syntactic positions, semantic roles, and clausal environments, models can accurately predict which entities are "alive" in a reader's mental model and which are destined to fade away.

The Syntactic Hierarchy: Grammatical Roles as Attention Cues

The most robust and widely utilized structural shortcut is the surface syntactic hierarchy. Across diverse languages, the grammatical role of a noun phrase serves as a direct proxy for its salience. The standard ranking follows a clear progression: Subject > Direct Object > Indirect/Oblique Object.

As Tetreault demonstrated in Analysis of Syntax-Based Pronoun Resolution Methods, simple algorithms operating purely over syntactic parse trees can outperform far more complex discourse-tracking models. The structural position of a noun phrase is inherently tied to its cognitive prominence.

This phenomenon manifests clearly in the "lifespan" of discourse entities. In The Life and Death of Discourse Entities: Identifying Singleton Mentions, Recasens et al. showed that entities introduced in the subject position or at the very beginning of a sentence are statistically much more likely to form long coreference chains, whereas entities in peripheral positions often act as "singletons" mentioned only once. Consequently, reference resolution systems achieve significant performance gains by prioritizing candidate antecedents whose previous mentions occupied prominent grammatical slots (Improving Pronoun Resolution by Incorporating Coreferential Information of Candidates).

This syntactic shortcut is not an artifact of English grammar; it appears to be a cross-linguistic universal. In a large-scale study using universal annotations, the authors observed that across ten diverse language families, the most likely antecedent for a pronoun is consistently found in a subject or object position (Investigating Multilingual Coreference Resolution by Universal Annotations). Even in informal or ungrammatical text, such as online news comments, the grammatical subject remains the most reliable heuristic for identifying the primary target of an opinion (Opinion Target Extraction in Chinese News Comments).

Crucially, surface syntax often trumps deep semantics when tracking immediate focus. Dahl argued in Focusing and Reference Resolution in PUNDIT that surface syntactic positions (such as direct objects and subjects) provide a more stable indicator of discourse focus than deep thematic roles, allowing systems to maintain focus tracking even when a sentence shifts from active to passive voice.

The Thematic Hierarchy: Deep Semantic Roles

While surface syntax is a powerful cue, it can be altered by stylistic choices. When surface cues are ambiguous or absent, language processing defaults to a deeper layer of organization: the thematic hierarchy. This hierarchy ranks semantic roles based on their inherent agency, typically prioritizing Agent > Patient/Theme > Goal/Source/Location.

This hierarchy is essential for resolving structural and scopal ambiguities. For instance, Poesio showed in Assigning a Semantic Scope to Operations that in the absence of explicit discourse context, human language processing defaults to a thematic hierarchy—prioritizing agents over themes—to determine the scope of operators (like "every" or "a").

In discourse tracking, Kong et al. demonstrated in Employing the Centering Theory in Pronoun Resolution from the Semantic Perspective that tracking the agents of primary verbs provides a significantly cleaner signal for the discourse center than tracking patients or recipients. Conversely, in natural language generation (NLG), leveraging a language-independent thematic hierarchy allows systems to map abstract semantic representations into grammatically correct word orders across multiple languages without requiring exhaustive, language-specific syntactic rules (Large scale language independent generation using thematic hierarchies).

The Clausal Hierarchy: Foregrounding vs. Backgrounding

Salience is also determined by the macro-structure of a sentence—specifically, how predicates are grouped into clauses. Main clauses serve to foreground information, while subordinate, relative, or appositive clauses background it.

This clausal hierarchy acts as a filter for attention. Upadhyay et al. leveraged this mechanism in "Making the News": Identifying Noteworthy Events in News Articles to isolate "news-pegs"—the core noteworthy events in news articles—by prioritizing events located in main clauses and active voice constructions over those buried in subordinate structures.

Similarly, the clausal hierarchy dictates how we track abstract concepts like time. Kameyama et al. demonstrated in Temporal Centering that time intervals introduced in main clauses are much stronger candidates for future temporal anchors than those in subordinate clauses. When tracking narrative perspective, the overall prominence of a character is often a competitive tug-of-war between global discourse factors (such as who was introduced first) and local clausal cues (who is the subject of the immediate main clause) (Coming in, or going out? Measuring the effect of discourse factors on perspective prominence).

Applications: Resolving the Unspoken and Generating the Natural

Understanding why structural hierarchies work allows us to build more elegant and efficient NLP systems. Instead of treating coreference or parsing as massive, unconstrained search problems, models can use structural salience to prune hypotheses.

Resolving Zero Anaphora

In "pro-drop" languages like Chinese and Japanese, speakers frequently omit pronouns entirely when the referent is obvious. Resolving these "zero pronouns" or implicit roles requires a precise model of what is currently salient (Casting Implicit Role Linking as an Anaphora Resolution Task).

Chen et al. showed in Chinese Zero Pronoun Resolution: An Unsupervised Probabilistic Model Rivaling Supervised Resolvers that syntactic context features, such as sentence distance and grammatical positioning, provide the most critical predictive power for resolution, outperforming semantic attributes. By modeling salience as a learnable ranking problem over grammatical case markers and predicate-argument structures, systems can dynamically maintain a cache of salient entities, drastically reducing the computational search space while maintaining high accuracy (Capturing Salience with a Trainable Cache Model for Zero-anaphora Resolution; Entity-Centric Joint Modeling of Japanese Coreference Resolution and Predicate Argument Structure Analysis).

Surface Realization in NLG

The inverse problem occurs in Natural Language Generation: given an abstract thought, how should a machine structure the sentence? Conklin et al. established in Salience: the Key to the Selection Problem in Natural Language Generation that the inherent salience of an object in a scene directly dictates its early placement in a natural language description.

To make generated text feel natural, systems must align logical prominence with psychological prominence (Deriving the Communicative Structure in Applied NLG). Zarrieß et al. demonstrated in Underspecifying and Predicting Voice for Surface Realisation Ranking that integrating discourse status (given vs. new) and animacy into a ranking model allows an NLG system to automatically choose the most natural grammatical voice (active vs. passive) and word order, ensuring that highly salient entities are consistently pushed to the front.

Alternative Perspectives: Information Structure

While grammatical and thematic hierarchies are powerful, some researchers argue that they are merely surface manifestations of a deeper underlying phenomenon: information structure.

Strube et al. proposed Functional Centering (Functional Centering), a framework that replaces rigid grammatical hierarchies entirely with information-theoretic rankings. They argue that an entity's prominence is determined by whether it represents context-bound given information or unbound new information.

In languages with highly flexible word order, like German or Danish, the physical position of a word—such as the "preverbal" slot in a main clause—acts as the primary structural signal for topicality, overriding standard grammatical roles (A Corpus-Based Approach to Topic in Danish dialog). However, whether framed as syntax, semantics, or information structure, the core insight remains the same: the structural organization of text is a deliberate map designed to guide and prioritize the reader's attention.

Key Foundational Work

The Use of Dynamically Extracted Context for Anaphoric Reference Resolution (1977) — contributed 1 concept elements to this direction.
Control of Inference: Role of Some Aspects of Discourse Structure-Centering (1981) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_syntactic_role_prioritization_e01b9d27,
  title = {syntactic role prioritization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/syntactic-role-prioritization/rdr_e01b9d27826e4e5c99d8e9809874e277},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_syntactic_role_prioritization_e01b9d27_v1,
  title = {syntactic role prioritization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/syntactic-role-prioritization/rdr_e01b9d27826e4e5c99d8e9809874e277/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake