Constraining Referent Accessibility through Hierarchical Discourse Structure

36 concepts24 papers1 proposal

When we read or listen to natural language, we effortlessly track entities across sentences. If a text mentions "a scientist," "a laboratory," and later uses the pronoun "she," our brains instantly map the pronoun back to the scientist, not the laboratory.

However, building computational systems to resolve these anaphoric links is notoriously difficult. A pronoun cannot simply link to any previously mentioned noun phrase. Instead, language operates under strict structural rules that dictate which entities are "visible" or accessible at any given point in a discourse.

To model this, researchers developed frameworks that treat discourse not as a flat sequence of words, but as a dynamic, hierarchical structure. By organizing discourse into nested logical boxes or tree-like graphs, we can geometrically define and constrain the search space for anaphora resolution.

The Geometry of Meaning: Discourse Representation Theory

The foundational framework for modeling this structural visibility is Hans Kamp’s Discourse Representation Theory (DRT). As Guenthner (1984) explains, DRT translates natural language sentences into formal logical structures called Discourse Representation Structures (DRSs).

Instead of evaluating sentences in isolation, a DRS acts as a mental model of the ongoing discourse, containing a set of discourse referents (the entities introduced) and conditions (the properties and relations ascribed to them). Guenthner et al. (1983) demonstrated that these structures provide a formal, rule-based method for defining accessibility relations between pronouns and potential antecedents.

The core intuition behind DRT's accessibility mechanism relies on a hierarchical system of nested structures, often visualized as "discourse boxes." As Frey et al. (1983) showed, when a text introduces logical operators like quantification, negation, or conditionals, it spawns a new, nested sub-box.

┌────────────────────────────────────────┐
│  x (The Scientist)                     │
│  scientist(x)                          │
│  ┌──────────────────────────────────┐  │
│  │  y (A Book)                      │  │
│  │  book(y)                         │  │
│  │  reads(x, y)                     │  │
│  └──────────────────────────────────┘  │
│  *Accessible here: x                   │
│  *Not Accessible here: y               │
└────────────────────────────────────────┘

An entity introduced inside a deeply nested box is bound by that local scope. A pronoun introduced later in the main, outer box cannot "reach inside" to link with it. For example, in the sentence "If a scientist reads a book, she enjoys it," the book is trapped inside the conditional sub-box. If a subsequent sentence says, "It was a bestseller," the pronoun "It" cannot legally access the book from the previous sentence because of these formal scoping rules.

To computationally implement this, Wada et al. (1986) mapped these nested boxes into a tree structure of discourse reference markers. By traversing this tree upward from the pronoun's location, an anaphora resolution algorithm can systematically filter out invalid candidates, enforcing grammatical and semantic agreement constraints along the way.

While DRT excels at tracking individual entities within sentences, discourse also exhibits a macro-structure. Sentences group together into segments (e.g., arguments, explanations, or narratives) that relate to one another hierarchically. This macro-structure introduces its own constraints on accessibility, particularly when pronouns refer to entire concepts or clauses rather than single nouns—a phenomenon known as discourse deixis.

Webber (1988) established that reference to previous discourse segments is strictly constrained by the structural position of those segments within a hierarchical discourse tree. Specifically, a speaker can only refer back to segments residing on the right frontier of the tree.

        [Segment 1: Main Topic]
               /       \
 [Segment 2: Detail]   [Segment 3: Elaboration]  <-- Right Frontier

The right frontier consists of two things: the most recently completed segment, and any higher-level parent segments that remain "open" for further elaboration. Once a sub-topic is closed and the speaker moves on to a new coordinate point in the conversation, the internal nodes of that closed sub-topic are sealed off. They are no longer accessible for immediate pronominal reference because they have been pushed off the active processing path.

Managing Expectations and Global Coherence

The right frontier is not entirely static; it dynamically contracts and expands based on linguistic cues. For instance, certain discourse markers create forward-looking expectations (e.g., "On the one hand..." signals that a matching "On the other hand..." is imminent).

Cristea et al. (1997) formalized this by introducing an inner frontier constraint. When a forward-looking expectation is active, it severely restricts where new information can attach in the discourse tree. By narrowing the search space, this mechanism ensures that the hierarchical structure remains consistent with the linear sequence of the text, preventing interpretations that would prematurely close a required open slot.

To scale this structural approach to long, complex texts, Cristea et al. (1998) developed Veins Theory. This theory computes sequential paths—called "veins"—through a discourse tree. A unit's vein expression identifies the exact subset of chronologically preceding units that are "visible" to it. By defining these precise global accessibility domains, Veins Theory extends local centering constraints to global discourse, significantly improving long-range anaphora resolution.

Applications Across Linguistic Domains

The insight that hierarchical structure dictates accessibility has proven highly adaptable, solving diverse natural language processing challenges:

  • Functional Definite Descriptions: Phrases like "the author" or "the top floor" behave like pronouns because they depend on a previously introduced anchor entity. Wada (1994) demonstrated that by treating these descriptions as pronominal-type discourse referents within a DRT framework, systems can resolve their relational conditions using standard accessibility rules.
  • Machine Translation: Translating pronouns between languages with different gender or honorific systems (like English to Japanese) requires deep contextual awareness. Wada (1990) solved this by appending a DRS directly to syntactic functional structures, ensuring that the translation engine can access the broader discourse context to select the correct target pronoun.
  • Punctuation as Structural Cues: Punctuation marks are more than grammatical ornaments; they signal how discourse segments attach to one another. The authors of An Information-Based Approach to Punctuation utilized Segmented Discourse Representation Theory (SDRT) to show how punctuation explicitly guides the construction of discourse trees, directly influencing abstract anaphora resolution by shaping the right frontier.
  • Complex Semantic Objects: Hierarchical accessibility frameworks have also been successfully extended to handle non-standard linguistic phenomena, such as tracking references to components of decomposable idioms (Fischer et al. 1996) and dynamically introduced locative prepositional phrases (Denand et al. 2004).

Ultimately, whether navigating the nested boxes of sentence-level semantics or the branching frontiers of global discourse, the core principle remains the same: language relies on geometry. By mapping text into hierarchical structures, we gain the mathematical constraints necessary to make sense of how ideas connect over time.

Key Foundational Work

Rules for Pronominalization (1983) — contributed 2 concept elements to this direction.
Automatic Construction of a Knowledge Base by Analysing Texts in Natural Language (1983) — contributed 2 concept elements to this direction.
Automatic Construction of Discourse Representation Structures (1984) — contributed 2 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_hierarchical_discourse_accessibility_4b1cb906,
  title = {hierarchical discourse accessibility},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hierarchical-discourse-accessibility/rdr_4b1cb90620ba4086a29c6a76ff99c357},
  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_hierarchical_discourse_accessibility_4b1cb906_v1,
  title = {hierarchical discourse accessibility},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hierarchical-discourse-accessibility/rdr_4b1cb90620ba4086a29c6a76ff99c357/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