Hierarchical Discourse Trees for Content Selection

52 concepts33 papers1 proposal

Traditional natural language processing often treats a document as a flat sequence—a "bag of words" or an ordered list of sentences. While superficial cues like word frequency or position can act as rough proxies for importance, they fundamentally lack structural awareness. As Annie Louis et al. demonstrate in Discourse indicators for content selection in summarization, relying solely on these surface features frequently leads to selecting sentences that are statistically common but rhetorically insignificant. To truly understand what matters in a text, we must look at how its arguments are built.

Text as a Hierarchical Skeleton

The core intuition behind hierarchical content selection is that human communication possesses an underlying logical architecture. We can model a document not as a flat string, but as a nested tree structure.

At the foundation of this perspective is Rhetorical Structure Theory (RST). As formalized by Daniel Marcu in The Rhetorical Parsing of Natural Language Texts, text can be broken down into Elementary Discourse Units (EDUs) connected by asymmetric relationships. Every relationship distinguishes between a nucleus—the core message essential to the narrative—and a satellite—the auxiliary information providing background, contrast, or elaboration.

If you remove a satellite, the text remains coherent; if you remove a nucleus, the local discourse collapses. This structural asymmetry provides an elegant, deterministic blueprint for content prioritization.

Why Tree Depth Predicts Salience

Why does this tree structure map so cleanly to human judgments of importance? The answer lies in the global topology of the discourse tree.

When a document is parsed into a global rhetorical tree, the most critical, overarching claims are promoted toward the root. Conversely, highly granular details, specific examples, and contextual qualifiers are pushed down into deeply nested sub-clauses.

Consequently, a unit's hierarchical depth serves as an inverse proxy for its salience. Research by Yuta Kikuchi et al. in Summarizing a Document by Trimming a Nested Tree Structure shows that units closer to the root consistently represent the core message, while deeper nodes contain peripheral details. In fact, Jun Suzuki et al. found that the accuracy of identifying just the root node is the single largest determinant of quality in tree-based summarization systems (Dependency-based Discourse Parser for Single-Document Summarization).

Algorithmic Exploitation of Discourse Trees

Knowing that text is a tree allows us to transform qualitative linguistic theories into quantitative optimization problems. Researchers exploit this hierarchy through several key mechanisms:

Tree Trimming and Knapsack Formulations

Rather than scoring sentences independently, tasks like summarization can be framed as a global optimization problem over trees. For instance, Tsutomu Hirao et al. in Single-Document Summarization as a Tree Knapsack Problem solve an Integer Linear Programming problem that explicitly penalizes text units based on their depth in the discourse hierarchy, ensuring global information is prioritized over local details.

Exponential Depth Decay

For tasks like document-level sentiment analysis, treating all clauses equally introduces noise from peripheral context. In Improving Sentiment Analysis with Document-Level Semantic Relationships from Rhetoric Discourse Structures, Joscha Märkle-Huß introduces a weighting model that applies exponential decay to the influence of text units based on their tree depth. This ensures that the primary structural skeleton dominates the final sentiment prediction, preventing auxiliary clauses from overshadowing the author's true intent—a technique mirrored by Parminder Bhatia et al. in Better Document-level Sentiment Analysis from RST Discourse Parsing.

Structural Inductive Biases in Neural Networks

In deep learning, discourse trees act as a powerful guide. Yangfeng Ji et al. utilize recursive neural networks with unnormalized attention mechanisms to naturally steer models toward central, summary-like components (Neural Discourse Structure for Text Categorization). More recently, Dongqi Liu 0001 et al. integrated rhetorical relation distributions directly into the Transformer architecture (Incorporating Distributions of Discourse Structure for Long Document Abstractive Summarization), allowing models to navigate long-context documents by following structural salience rather than simple token proximity.

Applications Across NLP

This structural prioritization generalizes across a wide array of language tasks:

Ultimately, hierarchical content selection works because it respects the way humans organize thought. By treating text as a logical tree, algorithms can look past the surface noise of word frequencies and directly target the structural core of a document.

Key Foundational Work

Abstract Generation Based on Rhetorical Structure Extraction (1994) — contributed 1 concept elements to this direction.
Generating Medical Case Reports with the Linguistic String Parser (1986) — 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_hierarchical_content_selection_fdeb93d3,
  title = {hierarchical content selection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hierarchical-content-selection/rdr_fdeb93d36a524ea58a62de02763cc0df},
  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_content_selection_fdeb93d3_v1,
  title = {hierarchical content selection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hierarchical-content-selection/rdr_fdeb93d36a524ea58a62de02763cc0df/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