On the Independence of Discourse Structure and Semantic Domain

Charlotte Linde, Joseph A. GoguenACL 1980 · DOI

Introduction

In traditional linguistics, the sentence was long considered the final frontier of structural analysis. However, when humans speak at length—whether giving a tour of an apartment, explaining a tax form, or planning a political strategy—their utterances are not merely a random string of sentences. They possess a high-level organization. A central debate in the field is whether this organization is a reflection of the "real world" (the semantic domain) or a property of language itself.

This paper argues that discourse structure is largely independent of the subject matter. By analyzing diverse datasets, the authors demonstrate that humans consistently "tree-ify" information, utilizing a predictable set of transformations to navigate complex ideas regardless of the topic at hand.

The Formal Model: Discourse as a Tree

The authors propose a formal model where discourse is viewed as a dynamic tree structure. In this framework, a conversation is built through a sequence of specific transformations on an initial empty tree:

  • Addition: The introduction of a new piece of information or a new node.
  • Deletion/Substitution: The removal or modification of ideas, a process particularly common in collaborative group planning.
  • Pointer Movement ("Pops"): Shifting the "focus of attention" back up the tree to a previous branching point to begin a new sub-topic.

To validate these structures, the researchers look for Text Marking—specific linguistic cues that signal a listener's position within the tree. Phrases like "On the other hand" or "Now as you're coming into..." serve as navigational markers for the listener.

Methodology and Datasets

The paper synthesizes findings from several distinct datasets to test the universality of these discourse rules:

  1. Apartment Layouts: 72 descriptions of New York apartments where speakers gave "tours" of their homes.
  2. Watergate Transcripts: Transcripts of President Nixon and his advisors engaged in high-stakes planning and evaluation.
  3. Explanations: Accounts of why individuals chose specific income tax forms or career paths.
  4. Morning Routines: An experimental dataset where participants described the process of getting their families ready in the morning.

Key Insights: The Universal Preference for Trees

The "Tree-ification" of Graphs

One of the most striking findings involves the description of spatial layouts. Many apartments contain "loops" (e.g., a circular path from the kitchen to the dining room to the hall and back to the kitchen). Mathematically, these layouts are graphs, not trees. However, the study found that 96% of speakers "cut" these loops in their descriptions to maintain a tree structure. This suggests a cognitive bias: humans find it significantly easier to process and produce tree-structured information than complex networks.

Depth-First Navigation

When faced with a branching path in a description, speakers almost universally adopt a "depth-first" strategy. They follow one branch to its logical conclusion before "popping" back to the original fork to begin the next branch. Only a tiny minority (roughly 4%) utilized a "Map" or breadth-first strategy.

The "And" Rule and Pre-marking

The most common way to branch a conversation is through the word "and." Because this is the default mode of expansion, it requires minimal marking. Conversely, for complex logical branches—such as "Either/Or" or "If/Then" scenarios—speakers almost always signal the structure before starting the first branch. This "header" helps the listener prepare for the mental mapping of a more difficult logical structure.

Social Dynamics and Tree Editing

In the context of group planning, such as the Watergate transcripts, the tree is not just built; it is edited. The authors observed that social activities like criticism or modification map directly onto tree transformations. When an advisor challenges a plan, they are effectively proposing a deletion or a substitution of a specific subtree within the discourse structure.

Conclusion and Future Directions

The findings suggest that despite the massive differences between describing a kitchen and planning a political cover-up, the underlying tree-transformation rules remain consistent. While the authors stop short of claiming a single, unified theory of discourse, they identify these "meta-rules" as universal constants across the domains investigated.

Future work may focus on "meta-systems"—computational models or AI designed to handle a variety of discourse types by utilizing these universal tree-building rules. However, the authors acknowledge that more domains must be investigated to formalize the rules for non-standard or "Map-based" descriptions which were underrepresented in the current study.

Summary updated

BibTeX
@inproceedings{linde1980independence,
  author = {Charlotte Linde and Joseph A. Goguen},
  title = {{On the Independence of Discourse Structure and Semantic Domain}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1980},
  doi = {10.3115/981436.981448},
}
Figure