Genre Variations in Discourse and Linguistic Signaling

36 concepts27 papers1 proposal

When we read a tweet, an academic paper, or a recipe, we intuitively switch our linguistic expectations. We do not expect a social media post to use the rigid logical transitions of a legal brief, nor do we expect an encyclopedia entry to use the emotional metaphors of a romance novel. Humans adapt their communication strategies completely based on the genre of the text.

However, natural language processing (NLP) models often treat text as a homogenous medium, assuming that structural, semantic, and grammatical signals operate under universal rules. A broad body of research demonstrates that this "one-size-fits-all" assumption fails. Linguistic signals—ranging from punctuation and pronouns to underlying logical connections—are deeply dependent on text genre. Understanding why and how these signals vary across genres is essential for building models that can truly comprehend and generate natural language.

The Structural Shift in Discourse Relations

How do we connect ideas? In text, the logical links between clauses (such as cause-and-effect, contrast, or temporal progression) can be explicit (using connectives like "because") or implicit (left for the reader to infer). Research shows that the distribution of these discourse relations is fundamentally governed by genre rather than language.

As shown by Sárka Zikánová et al. in their work on the DiscoGeM 2.0: A Parallel Corpus of English, German, French and Czech Implicit Discourse Relations, narrative literature consistently favors temporal relations across different languages, whereas argumentative texts prioritize logical links like reason and purpose. Similarly, Merel C. J. Scholman et al. established in DiscoGeM: A Crowdsourced Corpus of Genre-Mixed Implicit Discourse Relations that informative texts are heavily dominated by additive and descriptive relations. Even when a speaker and topic are held constant, Ines Rehbein demonstrated in On the role of discourse relations in persuasive texts that the underlying discourse relations shift depending on the communicative setting (e.g., articles versus interviews).

When we move to informal spaces like social media, these patterns shift dramatically. Tatjana Scheffler et al. observed in Annotating Shallow Discourse Relations in Twitter Conversations that Twitter conversations exhibit a much higher prevalence of contingency (cause-effect or conditional) relations and more fragmented argument structures compared to formal written text. An NLP model trained exclusively on news text will struggle to parse the rapid-fire, fragmented logic of social media because the underlying distribution of how ideas connect is entirely different.

Genre-Specific Signaling and Punctuation

Even when the underlying relation is the same, the surface-level cues used to signal it change. Yang Janet Liu investigated discourse signals across academic, how-to, interview, and news genres in Beyond The Wall Street Journal: Anchoring and Comparing Discourse Signals across Genres, discovering that many signals are highly genre-specific rather than universal. What serves as a clear structural signpost in one genre might be completely ambiguous or absent in another.

This variation extends down to orthography and punctuation. Punctuation marks are often dismissed as mere grammatical traffic signs, but Ming Yue demonstrated in Discursive Usage of Six Chinese Punctuation Marks that in Chinese news commentaries, specific punctuation marks act as reliable rhetorical signals that establish structural weight and link ideas. However, the author notes that these patterns are highly genre-dependent; the rhetorical conventions of a news commentary do not map cleanly onto fiction or social media.

Semantic and Lexical Adaptations

Genre also dictates how words are interpreted and selected. Consider metaphor detection: a verb that is almost always literal in casual conversation might be frequently metaphorical in academic writing. Beata Beigman Klebanov et al. showed in Semantic classifications for detection of verb metaphors that grouping verbs into semantic classes improves metaphor detection, but accurate detection requires genre-aware modeling because a word's literal-versus-metaphorical profile shifts with the communicative context.

Lexical choice and formulaic language are similarly constrained:

Grammatical and Deictic Variations

Beyond vocabulary, core grammatical choices like word order and pronoun usage adapt to genre constraints. For example, demonstrative expressions like "this" and "that" rely on context to point to things (deixis). Jeanette Gundel et al. proposed a framework in On the Generation and Interpretation of Demonstrative Expressions showing that the choice between these demonstratives and their grammatical forms is significantly influenced by discourse genre, meaning that anaphoric resolution algorithms must be genre-aware.

Similarly, Agnès Tutin and R. Kittredge highlighted in Lexical choice in context: generating procedural texts that procedural texts (like recipes) exhibit highly distinct anaphoric patterns that require domain-specific analysis during text generation. Even structural features like word order are bound by these constraints; Amanda Kann noted in Are Translated Texts Useful for Gradient Word Order Extraction? that word order extraction patterns are highly specific to the formal speech genre of the underlying corpus, warning that these patterns may not generalize to informal or literary text.

Why Genre-Aware Modeling Works

Why does incorporating genre information make models work better? If a model assumes a single global distribution for language, it faces massive ambiguity. A specific connective, punctuation mark, or clause type can mean entirely different things depending on whether it appears in an encyclopedia or a tweet.

[Generic Model] ---> Sees "Punctuation/Connective" ---> Guesses single global meaning (High Error)

                      /---> If Genre = News Commentary ---> Rhetorical Structure Signal
[Genre-Aware Model] -|
                      \---> If Genre = Social Media    ---> Casual Grammatical Marker

By forcing a single "correct" label across all text types, models fail to capture human-like interpretation. To solve this, Frances Yung et al. demonstrated in Label distributions help implicit discourse relation classification that training models on the distribution of human interpretations (rather than a single hard label) combined with explicit genre tokens significantly improves implicit discourse relation classification.

The genre token acts as a statistical prior, shifting the model's internal probability distribution to match the specific rhetorical conventions of the target text type. As highlighted by Amir Zeldes et al. in Introduction to Discourse Relation Parsing and Treebanking (DISRPT): 7th Workshop on Rhetorical Structure Theory and Related Formalisms and Deniz Zeyrek and Kezban Başıbüyük in TCL - a Lexicon of Turkish Discourse Connectives, moving away from generic text processing toward genre-specific sense distributions and functional text units—such as those explored by Rachele Sprugnoli et al. in The Content Types Dataset: a New Resource to Explore Semantic and Functional Characteristics of Texts—is the key to unlocking robust, context-aware natural language understanding.

Key Foundational Work

On the Generation and Interpretation of Demonstrative Expressions (1988) — contributed 3 concept elements to this direction.
A Contrastive Analysis of the Use of Definite Articles in English Scientific Texts and in English Literature (1982) — 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_genre_dependent_demonstrative_usage_513dc116,
  title = {genre-dependent demonstrative usage},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/genre-dependent-demonstrative-usage/rdr_513dc1165c0c48e881631362b5c356b7},
  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_genre_dependent_demonstrative_usage_513dc116_v1,
  title = {genre-dependent demonstrative usage},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/genre-dependent-demonstrative-usage/rdr_513dc1165c0c48e881631362b5c356b7/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