Why Discourse Models Fail to Generalize Across Genres

128 concepts106 papers1 proposal

When we read a text, we do not just process a sequence of independent sentences. We naturally track a high-level structural blueprint: an introduction sets the stage, a contrastive marker like "however" shifts the direction, and a causal phrase links an action to its consequence. Discourse modeling aims to teach machines this structural grammar, enabling them to understand the overarching logical flow of documents.

However, a persistent bottleneck plagues the field: a model that excels at mapping the logical flow of a newspaper article often fails completely when applied to a legal brief, a scientific abstract, or a transcript of a TED talk. This cross-domain performance drop reveals that our models are not learning a universal logic of text organization. Instead, they are deeply tethered to the specific rhetorical conventions of the domains on which they were trained.

The Illusion of Universal Coherence

Why is transferring discourse knowledge so difficult? Intuitively, one might assume that the fundamental principles of logical coherence—such as cause-and-effect or temporal progression—are universal across human communication. If a model understands how sentences link together in one context, it should understand them in another.

In reality, models often rely on domain-specific statistical patterns and surface-level lexical cues rather than abstract logical reasoning. In Evaluating Document Coherence Modelling, Shen et al. demonstrated this by testing models on an "intruder sentence" detection task. While state-of-the-art language models performed exceptionally well on familiar topics, their performance dropped sharply when introduced to new domains. The authors concluded that the models were largely memorizing domain-specific facts and shallow text patterns rather than internalizing a generalized discourse logic.

This domain gap can be even more disruptive than a language barrier. For instance, Eger et al. found in Cross-lingual Argumentation Mining: Machine Translation (and a bit of Projection) is All You Need! that bridging the gap between different text genres (such as formal essays versus informal reviews) posed a significantly greater challenge for argumentation mining models than translating text across entirely different languages.

Rhetorical Fingerprints and Structural Variance

The root of the problem lies in how heavily the distribution and expression of discourse relations vary by genre. Different types of text use completely different structural building blocks.

Consider the difference between written journalism and spoken communication. News articles are heavily dominated by expansion and elaboration relations—sentences designed to unpack details about a central event. Conversely, as shown by Long et al. in TED-CDB: A Large-Scale Chinese Discourse Relation Dataset on TED Talks, spoken monologues like presentations rely far more on contingency and temporal links to guide an audience through a live narrative. A discourse parser trained exclusively on news text will inevitably misinterpret or overlook these conversational transitions.

These patterns are so distinct that the linear sequence of rhetorical relations can serve as a structural signature. Knaebel et al. explored this concept in Discourse Sense Flows: Modelling the Rhetorical Style of Documents across Various Domains, demonstrating that the specific ordering of rhetorical moves acts as a "fingerprint" capable of automatically distinguishing between genres like political speeches, scientific abstracts, and news.

Furthermore, the underlying "situation entities"—the events, states, and generic statements that form a text—change drastically across prose styles. Palmer et al. noted in A Sequencing Model for Situation Entity Classification that models trained on structured news data suffer severe performance drops when applied to popular lore or creative writing, because different discourse modes possess unique, non-transferable sequencing signatures.

The Granularity Gap in Downstream Tasks

When a discourse model fails to generalize, it creates a cascading failure in downstream applications like text simplification, sentence insertion, or summarization.

Many genres rely on highly specialized "move structures" that generic discourse frameworks fail to capture. For example, in news simplification, a model must distinguish between a main event and an anecdotal aside. As Zhang et al. argued in Predicting Sentence Deletions for Text Simplification Using a Functional Discourse Structure, generic rhetorical structures lack the granularity required to identify these specific functional roles, preventing models from applying genre-specific deletion logic.

Similarly, tasks that require structural reasoning, such as placing a stray sentence back into a paragraph, are highly sensitive to document density and internal logic. Wu et al. demonstrated in InsertGNN: A Hierarchical Graph Neural Network for the TOEFL Sentence Insertion Problem that models trained on condensed scientific abstracts struggle when transitioning to expanded narrative passages, because the underlying expansion styles and rhetorical densities are fundamentally mismatched.

Moving Toward Domain-Agnostic Discourse

To overcome these limitations, the research community is shifting away from rigid, single-domain rules and moving toward architectures and datasets designed explicitly for cross-genre robustness.

One approach focuses on developing parsing frameworks that learn generalized principles of information organization. In Towards Domain-Independent Text Structuring Trainable on Large Discourse Treebanks, Guz et al. introduced a model designed to order and organize text units into discourse trees without relying on hard-coded, domain-specific rules. Concurrently, researchers are leveraging stronger underlying representations; Kobayashi et al. showed in A Simple and Strong Baseline for End-to-End Neural RST-style Discourse Parsing that scaling the pre-trained language model baseline can mitigate some out-of-distribution degradation, though the underlying domain sensitivity remains a core challenge.

Ultimately, solving the generalization problem requires training and evaluating models on data that reflects the true diversity of human discourse. In DiscoGeM: A Crowdsourced Corpus of Genre-Mixed Implicit Discourse Relations, Scholman introduced a crowdsourced corpus spanning literature, political speeches, and encyclopedic texts. By capturing the inherent ambiguity and varied distributions of logical relations across multiple communicative contexts, such efforts provide the foundational benchmarks needed to move from domain-specific pattern matching to a truly generalized understanding of document structure.

Key Foundational Work

The Use of Syntactic Clues in Discourse Processing (1985) — 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_cross_domain_discourse_generalization_2d57608a,
  title = {cross-domain discourse generalization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-domain-discourse-generalization/rdr_2d57608a982440b7ba706b13d9c969a4},
  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_cross_domain_discourse_generalization_2d57608a_v1,
  title = {cross-domain discourse generalization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-domain-discourse-generalization/rdr_2d57608a982440b7ba706b13d9c969a4/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