NFAF: Narrative Flattening Analysis Framework (0.5.0)

Author: Johennie Helton

Status of This Document: Working Draft Specification (v0.5.0)

Date: September 2025

Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)

Copyright (c) 2025 Johennie Helton

Abstract

The Narrative Flattening Analysis Framework (NFAF) provides a conceptual framework for identifying and evaluating the loss of narrative fidelity that can occur during AI-mediated transformation. It defines the dimensions of narrative flattening, establishes a shared vocabulary, and serves as the foundation for future empirical evaluation.

1. Status of This Document

This document is a Working Draft Specification of the Narrative Flattening Analysis Framework (NFAF). It is released for academic and research purposes and may be updated, replaced, or obsoleted at any time prior to version 1.0.

This specification addresses AI-mediated narrative flattening. Domain-emergent flattening is acknowledged but out of scope (see §10).

NFAF v0.5 presents a systematic framework derived from theoretical foundations and observed patterns in AI-mediated transformations. Broader empirical validation across corpora, languages, and domains, as well as eventual systematic implementations, remain priorities for future versions.

Purpose

• Defines terminology, analytic stages, and rubric dimensions for detecting narrative flattening in AI-mediated transformations (translation, summarization, rewriting, and related processes).
• NFAF is non-prescriptive: it separates detection from judgment.
• Intended for use in research, teaching, and non-commercial prototyping.
• White papers, case studies, and domain-specific applications SHOULD cite this specification but MAY extend it with illustrative examples.

Conformance An implementation conforms to this specification if it:

1. Applies at least one rubric dimension (§7) to a source/target narrative pair.
2. For conformance with this specification, implementers MUST carry out Stages 1–5 and 7 as defined in §6, adapting them as needed for their domain. Stage 6 is REQUIRED only in high-stakes contexts.
3. Produces an alignment manifest (§6.3) recording omissions, additions, or drift.
4. Discloses observed flattening indicators (§5.9).

Partial conformance MAY be declared when only a subset of stages or rubric dimensions is implemented, provided the deviation is explicitly documented.

Keywords MUST, MUST NOT, SHOULD, SHOULD NOT, and MAY are to be interpreted as described in [RFC 2119]. These directives specify how analysts should apply the framework consistently, not whether observed changes are desirable or problematic, such evaluations remain domain-specific.

Intellectual Property Notice NFAF is published as the pre-existing intellectual property of Johennie Helton. This specification is provided for research and citation purposes only.

• No commercial implementation rights or co-ownership are granted. Any commercial use or integration REQUIRES prior written permission from the author.
• For non-commercial research prototypes, users MUST cite this specification and MUST disclose any deviations.
• Redistribution SHOULD preserve version information (v0.5.0) and persistent identifiers (DOI, ORCID).

2. Citation

If you reference this framework in research or teaching, please cite as follows:

Helton, J. (2025). NFAF: Narrative Flattening Analysis Framework (v0.5.0). Zenodo. https://doi.org/10.5281/zenodo.16885076

The example above is formatted in APA style. Researchers MAY adapt to MLA, Chicago, or other disciplinary standards, provided the DOI is preserved.

Persistent Identifiers

DOI: 10.5281/zenodo.16885076
ORCID: 0009-0003-2175-3239

3. Abstract

NFAF (Narrative Flattening Analysis Framework) is a modular method for detecting and characterizing narrative flattening across AI-mediated transformations, including translation, summarization, rewriting, and related processes. Narrative flattening refers to the convergence of AI-generated text toward standardized patterns that affect stylistic voice, idiomatic texture, and cultural specificity. It manifests as erosion (loss of distinctive semantic, stylistic, cultural, or affective features) or inflation (addition of generic elements that override the original signal). A central mechanism is the hidden pivot: an implicit or explicit relay language through which AI systems may route mediations, with outputs that may show signs consistent with hidden-pivot processing. Narrative flattening is treated here as an empirical trace of hidden pivot effects, visible in measurable erosion of cadence, idiom, and cultural register. NFAF defines analytic stages and rubric dimensions that provide a reproducible basis for documenting and comparing flattening across corpora and domains. Grounded in narratology, translation theory, digital humanities, and NLP, NFAF offers a replicable way to analyze whether, where, and how AI outputs converge toward hidden pivots, and what nuance is lost in the process.

Table 1. Categories of narrative flattening and their indicators. Canonical pull = tendency of AI systems to gravitate toward standardized or widely circulated formulations, overriding text-specific variation.

This specification defines flattening in two distinct contexts:

• (a) AI-mediated flattening, the primary focus of NFAF, where narrative distinctiveness is eroded through translation, summarization, rewriting, or related processes performed by artificial intelligence

• (b) domain-emergent flattening, recognized but out of scope here, where flattening arises from practices internal to specific fields (e.g., journalistic false balance, legal compression of testimony, political simplification). This document addresses (a). For discussion of (b), see §10.

4. Introduction

The Narrative Flattening Analysis Framework (NFAF) provides a modular framework to detect, describe, and compare the loss, convergence, or homogenization of cultural, emotional, and stylistic features if present in narratives transformed by machine mediation such as translation, summarization, rewriting, or other processes.

NFAF is not domain-specific. It contains a common theoretical and methodological core while deliberately avoiding prescriptive thresholds or universal acceptability criteria. A reduction in variance that signals voice loss in literature, for example, may constitute acceptable compression in legal summarization. Interpretation must therefore be guided by domain expertise, use case requirements, and stakeholder values. For example, literary analysis may prioritize voice preservation and cultural authenticity; legal applications may emphasize semantic precision and structural fidelity; journalism may balance accessibility with source integrity. This modularity enables specialized implementations (e.g., Literature, Law) while maintaining consistent analytical stages and ethical principles.

NFAF recognizes two sources of flattening: those originating from AI mediation and those arising from domain emergence. NFAF applies to AI-mediated flattening, where narrative distinctiveness is eroded through artificial intelligence processes. Domain-emergent flattening, where flattening arises from professional practices themselves (e.g., journalistic false balance, legal compression of testimony, political simplification) is out of scope for v0.5.0 but is addressed as a possible extension in §10.

AI-mediated transformations often reflect algorithmic optimization toward pivot-language outputs that may appear universally fluent but are less contextually situated. NFAF positions these effects as systematic convergence patterns observable in outputs, regardless of the internal architecture of the model. NFAF treats such convergence as an empirical trace of hidden pivot effects, consistent with relay-language processing, whether implicit in training distributions or explicit in chained mediations.

Comparable critiques of homogenization can be found across fields. In translation studies, (Berman 2000) catalogued systematic distortions in his negative analytic, (Venuti 1995) emphasized the risks of domestication, and (Spivak 2000) highlighted the ethical duty to preserve singular voice. Recent work on machine translation reinforces these concerns: (Schneider 2022) identifies English functioning as a hidden pivot in digital multilingualism, and (Vieira 2023) highlights how infrastructures shape translation practices and reinforce convergence. In digital humanities and media theory, (Drucker 2014) critiques assumptions of interpretive neutrality, echoing concerns about voice and context.

NFAF complements rather than replaces quantitative metrics such as BLEU or ROUGE. It asks different questions entirely, focusing on dimensions of meaning preservation that surface-level similarity measures cannot capture. However, NFAF does not claim universal applicability: its categories derive primarily from Western narratology, translation studies, and journalistic and legal traditions, and may need extension or adaptation in diverse cultural and disciplinary contexts.

The following sections define NFAF’s key concepts and analytic dimensions, outlining the indicators and stages that make flattening measurable across different domains and mediation types.

Current focus: Applying NFAF to narrative voice in literary text, with emphasis on preserving cultural and emotional signals in AI-mediated transformation. Sections §7–§8 are illustrative; only §7 provides worked evidence in v0.5.0. The framework is grounded in both theoretical and observed patterns; broader empirical validation across corpora, languages, and domains, as well as potential systematic implementations, remain priorities for future versions.

Note: NFAF is published as pre-existing intellectual property of Johennie Helton. This specification is provided for research and citation purposes only. No commercial implementation rights or co-ownership are granted. For non-commercial research prototypes, cite this spec and disclose deviations.

5. Definitions

Note: Keywords MUST, MUST NOT, SHOULD, SHOULD NOT, and MAY are to be interpreted as described in [RFC 2119]. These directives specify how analysts should apply the framework consistently, not whether observed changes are desirable or problematic, such evaluations remain domain-specific.

NFAF analyzes output patterns, not internal mechanisms. When we describe 'pivots' or 'convergence effects,' we refer to observable regularities in transformed texts, not claims about how AI systems internally process information.

5.1 Narrative (adapts)

A narrative is any structured account of events, experiences, or arguments (whether linguistic, visual, or multimodal) that conveys meaning through deliberate sequencing, framing, or voice. This definition adapts broad narratological accounts (Genette 1980; Bal 1985; Ryan 2004), treating discourse and framing as central to meaning.

• Narratives MUST be analyzed with attention to distinctive signals such as voice, style, cultural cues.
• Narratives MAY be literary, journalistic, scientific, legal, cultural, or digital-born in form.
• Narratives MAY be linear or nonlinear, explicit or implicit, but are always shaped by formal, stylistic, and perspectival choices tied to context.
• Any mediation or transformation of a narrative MUST respect the integrity of these signals. The source SHOULD remain accessible alongside the modified version to prevent erasure of voice or context.

In this specification, narrative is considered primarily in the context of AI-mediated transformations (§6). Other forms of narrative reduction or simplification are acknowledged but treated separately under domain-emergent flattening (§10).

5.2 Narrative Flattening (defines)

Narrative flattening is defined by this specification as the reduction, erosion, or homogenization of a narrative’s distinctive signals (its voice, style, cultural cues, or emotional intensity) through processes of transformation or mediation. Note that when texts go through multiple AI mediations, narrative flattening effects might compound.

Comparable critiques are found in translation theory (Berman 1985; Venuti 1995; Spivak 2000), where mediation reshapes voice and culture. NFAF reframes these concerns as reproducible, algorithmic effects of AI mediation, echoing computational linguistics on semantic shift (Hamilton et al. 2016) and digital humanities critiques of interpretive neutrality (Drucker 2014).

Whether such reduction constitutes a problem depends on domain context and transformation goals; NFAF provides tools for detection and categorization, not evaluation.

5.3 Flattening

Flattening is the systematic reduction of variance in narrative voice across transformations and can be observed in several recurrent forms. Forms of flattening describe recurrent ways narrative distinctiveness is reduced during mediation:

• Erosion: Loss or weakening of key narrative elements, such as semantic nuance, stylistic rhythm, cultural references, or affective tone.
• Inflation: Insertion or exaggeration of generic features absent from the source.
• Speculative Addition/Inference: The introduction of details absent from the source, often drawn from the training data on canonical texts, precedent cases, or analogous contexts. This can prematurely resolve ambiguities in a way not supported by the original source.
• Omission/Compression (context-dependent): Removal or condensation of source segments (sentences, paragraphs, dialogue, quotations). In summarization, omission may be expected but should always be signaled to preserve accountability; in translation or rewriting, omission of substantive content constitutes flattening unless explicitly documented.

Whether by erosion, inflation, inference, or omission, the result is outputs that are more standardized, less contextually situated, and ultimately less distinctive at the cost of human voice and cultural nuance. The effects appear across domains: in literature as homogenized rhythm or reduced figurative density; in journalism as genericized news summaries; in law as legally imprecise translations.

5.3.1 Contexts of flattening

NFAF distinguishes two contexts:

• AI-mediated flattening (normative focus):
  Reduction of distinctiveness caused by artificial intelligence systems during translation, summarization, rewriting, or related transformations.

  • Analysts MUST evaluate AI outputs against the source narrative for losses of cadence, voice, cultural register, and other rubric dimensions (§7).
  • AI-mediated flattening MUST be documented using the analytic stages (§6).
  • Results SHOULD disclose both explicit omissions and subtle homogenizations.

• Domain-emergent flattening (acknowledged, out of scope in v0.5.0):
  Reduction of distinctiveness arising from institutional or professional practices (e.g., journalistic false balance, legal compression of testimony, political simplification).

  • Domain-emergent flattening is recognized but MUST NOT be treated as normative in v0.5.0.
  • Future extensions (e.g., NFAF-J for Journalism, NFAF-L for Law, NFAF-P for Politics) MAY adapt the framework to these contexts (§10).

Scope of v0.5.0
Unless explicitly stated otherwise, narrative flattening in this specification refers to AI-mediated flattening.

5.4 Pivot

In NFAF, a pivot is any intermediate form: linguistic, cultural, or semiotic, through which mediated processes converge.

NFAF distinguishes two key dimensions:

• Centroid Pivot (§5.4.2): the generalized phenomenon of convergence toward dominant defaults. Within this, hidden pivot (§5.4.1) designates cases where convergence toward an unobserved intermediary is inferred from outputs. Hidden pivot may manifest in both relay (multi-step) and single-step mediation.
• Pivot Form (§5.4.2): the specific centroid defaults (language, culture, register) that shape mediation outcomes.

Pivot (umbrella category)
├── Centroid Pivot = generalized convergence toward dominant defaults
├    └── Hidden Pivot = inferred convergence toward an unobserved intermediary
├    ├── Relay (multi-step mediation)
├    └── Single-step (direct mediation with inferred centroid pull)
└── Pivot Form = the specific centroid defaults (language + culture + register) shaping mediation

Figure 1. Pivot hierarchy.

Adaptation note: This extends the notion of a pivot language in translation studies and NLP, generalizing it to multimodal, cross-domain contexts and explicitly including cultural register.

    Input (language or modality)
          ↓
       [Pivot stage: single-step or relay]
          ↓
    Pivot Form (linguistic + cultural attractor)
          ↓
    Output (language or modality)

* Illustrative schematic only; not a claim about system internals.

Figure 2. General pivot schematic.

5.4.1 Hidden Pivot (adapts)

The term hidden pivot originates in translation studies, where English often functions as an invisible relay in multilingual systems (Schneider 2022; Vieira 2023). In NFAF, a hidden pivot is defined more broadly as a subtype of centroid pivot, identified when convergence toward an unobserved intermediary can be inferred from outputs.

In NFAF, hidden pivot operates as an interpretive lens for diagnosing convergence patterns, without presupposing a single causal mechanism. Such patterns may arise from relay processing, training data dominated by English, architectural constraints favoring English-like structures, or algorithmic choices (e.g., English as development and evaluation default).

Whereas in translation studies hidden pivot typically denotes a processing mechanism, in NFAF it denotes an output pattern: a way of reading systematic convergence without requiring knowledge of its source. The practical effect is that outputs bear the texture and influence of a centroid pull toward a pivot form: in many cases, English language and globalized cultural defaults.

Hidden pivots may appear in two forms: single-step and relay.

• A single-step pivot occurs when only the final output is visible but carries traces of an intermediate attractor.
• A relay pivot occurs when intermediate outputs reveal the pivot explicitly.

Empirical studies provide evidence that non-English to non-English translations often carry Anglicized traces, consistent with hidden pivot effects:

• FR→ES (Vanmassenhove & Shterionov 2019): Neural MT outputs exhibit reduced lexical richness, shorter clauses, and neutralized register, features aligned with English stylistic defaults, despite English not being source or target.
• FR→PT, ES→PT (Torres & Ramos 2018): Portuguese translations display overuse of Anglicized terminology and phraseology in contexts where Romance-native equivalents exist.
• IT↔ES, FR↔ES, PT↔FR (Popović et al. 2016): Documented “translationese” effects of simplification and normalization consistent with English-driven training bias.

These findings support the inference that English operates as a hidden pivot. NFAF treats such convergence as an empirical trace of hidden pivot processing: detectable in outputs even when English is neither the input nor the output language.

5.4.2 Centroid Pivot (defines; generalizes)

Centroid Pivot (CP) is a conceptual model coined in this specification to generalize the hidden pivot beyond multilingual translation, modeling convergence toward central forms in any modality. That is, CP denotes the systematic convergence of mediated outputs toward a culturally or statistically central form, regardless of domain or modality.

• Hidden pivots are a specific linguistic manifestation of CP.
• CP applies equally to text, visual, musical or other semiotic systems, where convergence toward dominant defaults can be observed.
• CP is not a claim about model architecture; it is an interpretive framework derived from regularities in outputs.
• CP MUST NOT be assumed uniform: different systems may exhibit different convergence patterns, depending on training distributions.

5.4.3 Pivot Form (defines)

A Pivot Form is the specific language + cultural register that functions as the attractor in a given mediation.

• Pivot Forms are training-data dependent and system-specific.
• They may differ across models.
• In relay scenarios, each system may apply its own Pivot Form, and may produce layered or compounded flattening.

   Example: P1 = Internet English + globalized culture; P2 = Mandarin + Chinese digital culture.

    Model A (Pivot Form P1) FR → [P1] → PT 
	+ 
	Model B (Pivot Form P2) PT → [P2] → ES
	
- Model A aligns FR to Pivot Form P1 (Internet English + culture).
- Model B aligns PT to Pivot Form P2 (Mandarin + Chinese culture).
- ES shows traces of both P1 and P2.

* Relay pivots illustrate system-specific convergence; variation reflects training data differences, not uniform architecture.

Figure 3. Multi-system relay pivot (distinct Pivot Forms).

Directives

• Analysts SHOULD document observed Pivot Forms descriptively (e.g., ‘Internet English + globalized culture’), not as fixed taxonomies.
• Analysts MUST treat pivot effects as output-based inferences, not assumptions about system internals.
• Identification of pivots MAY rely on observable regularities, whether or not intermediate outputs are visible.
• Pivot Forms MUST NOT be assumed neutral or universal; they may reflect systemic pulls toward dominant defaults.
• CP MUST NOT be assumed identical across systems; variation is expected and should be documented.

5.5 Voice & Cadence Metrics (adapts)

Voice & Cadence Metrics are quantitative and qualitative measures used to assess a text’s stylistic rhythm, pacing, and intensity, specifically to determine whether its narrative rises and falls are preserved during transformation. This definition adapts methods from stylistics (Leech & Short 1981), computational authorship attribution (Burrows 2002), and stylometric analysis (Stamatatos 2009) to the context of AI-mediated transformation.

• Metrics are descriptive indicators of variation and distinctiveness; they do not prescribe universal thresholds of acceptability.
• Because cadence conveys meaning, any reduction in rhythm, figurative density, or stylistic modulation MUST be recognized as flattening rather than “clarification.”
• Analysts MUST document cadence shifts observed across aligned segments (§6).
• Metrics MAY include:
  • Structural indicators: sentence length variance, clause complexity (e.g., prose, journalism).
  • Stylistic indicators: figurative density, idiomatic clusters (e.g., metaphor clusters in literary style).
  • Rhythmic indicators: punctuation-driven pacing, repetition, emphasis (e.g., legal clauses, journalistic headlines).
  • Dialogic/prosodic indicators: turn-taking, speech markers (e.g., emphasis, turn-taking in poetry or conversation).

Voice & Cadence Metrics provide the primary diagnostic layer for detecting loss of stylistic distinctiveness. They are descriptive of surface outputs, not claims about system internals; they can be observed and documented even when underlying model pathways remain unknown.

5.6 Drift (adapts)

Drift, in NFAF, refers to the systematic shifts—whether intentional or emergent, that arise when a narrative undergoes transformation. Drift alters either the scaffold (structure and discourse relations) or the semantics (meaning and resonance).

Within NFAF, drift is a descriptive category: it captures shifts in scaffold or semantics introduced by mediation. Some drifts contribute to flattening (§5.3), while others reshape structure or meaning without necessarily diminishing voice.

Drifts may be identified as structural or semantic; however, most transformations can be examined through both lenses. Researchers are encouraged to adopt the lens most relevant to their question and corpus.

5.6.1 Structural Drift

Structural Drift refers to changes in the narrative’s scaffold: the arrangement of segments, discourse relations (contrast, cause, elaboration), speaker turns, or quoted material.

• Analysts MUST document structural drift whenever scaffolding is altered without disclosure.
• Structural drift MAY manifest as:
  • Segment Omission: Missing dialogue, attributions, or text (null alignment).
  • Dialogue Flattening: Removal of speaker attributions or turns.
  • Event Collapsing: Fusion of steps that are meaningful to the narrative.
  • Inflation (structural): Addition of dialogue or attributions that alter discourse framing.

This category resonates with work on discourse structures and relations (Mann & Thompson 1988; van Dijk 2008).

5.6.2 Semantic Drift

Semantic Drift refers to systematic or cumulative changes in meaning between a source and transformed text. Surface fidelity alone is insufficient; analysts MUST document hidden shifts so that ambiguity and nuance are not silently overwritten.

• Semantic drift MAY manifest as:
  • Lexical Drift: Substitution, loss, or generalization of specific word choices.
  • Syntactic Drift: Shifts in sentence structure that alter emphasis or rhythm.
  • Pragmatic Drift: Changes to implied meaning, tone, or cultural resonance.
  • Inflation (semantic): Introduction of new propositional content absent from the source.

This notion aligns with linguistic theories of semantic change and with Digital Humanities accounts of interpretive “drift” in computational reading (Ramsay 2011; Jockers 2013). NFAF applies it specifically to shifts introduced by AI mediation.

5.6.3 Taxonomy Diagram

                 ┌───────────────┐
                 │     Drift     │
                 └───────┬───────┘
                         │
         ┌───────────────┴───────────────┐
         │                               │
 ┌───────▼───────┐               ┌───────▼───────┐
 │ Structural    │               │ Semantic      │
 │ Drift         │               │ Drift         │
 │ (scaffold)    │               │ (meaning)     │
 └───────┬───────┘               └───────┬───────┘
         │                               │
  - Segment omission              - Lexical drift
  - Dialogue flattening           - Syntactic drift
  - Event collapsing              - Pragmatic drift
  - Inflation (structural)        - Inflation (semantic)  

Figure 4 Taxonomy of drift in NFAF (structural vs semantic).

Normative scope: Both structural and semantic drift compromise narrative fidelity. Analysts MUST distinguish between them: structural drift alters scaffold and discourse relations, while semantic drift alters meaning and resonance.

5.7 Narrative alignment (adapts)

Narrative alignment is adapted in NFAF to mean the segment- or unit-level mapping between a source narrative and its transformed version. Alignment enables side-by-side analysis and is a prerequisite for identifying omissions, additions, or shifts in meaning.

• Analysts MUST perform alignment when evaluating AI-mediated transformations. Without such mapping, flattening may occur invisibly.
• Alignment MUST preserve transparent traceability between source and output segments, thereby maintaining the integrity of the source narrative.
• Alignment MAY be implemented using manual annotation, computational tools, or hybrid methods, provided the mapping is explicit and reproducible.
• Comparable practices include parallel corpora in translation studies, alignment tools in NLP (Gale & Church 1993), and text comparison methods in law, journalism, and scientific abstracts.

In NFAF, alignment is not only a methodological step but a humanistic accountability tool. It ensures that mediated texts remain anchored to their originals, preventing silent erasure of voice or context.

    Source Narrative     →     Transformed Narrative
    [Segment A] --------→ [Segment A’]
    [Segment B] --------→ [Segment B’]
    [Segment C] --------→ (omitted)
    [Segment D] --------→ [Segment D + added material]
    
    Alignment reveals omissions, additions, and shifts

Figure 5. Alignment schematic. Alignment exposes divergences between source and transformed narratives.

5.8 Signal Preservation (adapts)

Signal preservation is adapted in NFAF to mean the degree to which a narrative’s core elements - tone, imagery, prosody, and culturally specific forms - are retained during transformation. Preservation is a normative requirement: losses MUST be made visible and attributable to the mediation process.

• Analysts MUST disclose any loss of narrative signals so that readers can evaluate the meaning and integrity of the transformed text relative to its source.
• Preservation MUST include both stylistic signals (e.g., tone, imagery, prosody, cultural registers) and structural signals (e.g., speaker turns, quotations, paragraphing).
• The undocumented loss of these signals MUST be treated as flattening.
• Preservation MAY be partial, but omissions or compressions MUST be documented as such.

Signal preservation applies across domains, whether in literary rewriting, journalistic summarization, legal translation, or scientific abstraction. Without explicit preservation, flattening occurs invisibly, erasing both context and accountability.

Comparable concerns appear in archival science (provenance, integrity) and information theory, but NFAF repositions them around narrative voice, cultural register, and emotional resonance.

5.9 Flattening Indicators (defines)

Flattening indicators are defined in NFAF as reusable signs that a narrative’s distinctiveness has been reduced during an AI-mediated transformation. Their presence serves as a red flag, signaling potential erasure or overwriting of cultural and emotional meaning, and may necessitate human review.

• Indicators are observable symptoms.
• Drift (§5.6) names types of change; indicators operationalize these shifts for detection.
• Analysts MUST document flattening indicators whenever they appear.
• Indicators SHOULD be evaluated in relation to alignment (§5.7) and signal preservation (§5.8).
• NFAF should be understood as a located framework, reflecting particular theoretical traditions. Its categories may not map cleanly onto all cultural or narrative systems.

5.9.1 Qualitative indicators

5.9.1.1 Structural & Alignment Indicators

• Structural drift
• Segment omission: unmatched source segments (sentences/paragraphs) in alignment.
• Dialogue flattening: removal of quoted speech or speaker attributions; monologizing multi-turn scenes, i.e., reducing polyphonic scenes to a single speaker voice.
• Event collapsing: fusing discrete actions or time steps into generic summaries.
• Discourse link collapse: loss of explicit connectives (“however,” “therefore”), erasing contrast or causality.

5.9.1.2 Cadence Indicators

• Homogenized cadence: loss of rhythm, irony, suspense, or lyrical quality.
• Loss of prosodic markers/flattened pacing: reduced variance in sentence length or paragraph flow; disappearance of punctuation or figurative clusters that carry rhythm.

5.9.1.3 Stylistic Indicators

• Loss of idiomatic expressions: unique idioms replaced with generic phrases.
• Quotation loss / paraphrase masking: quotes replaced with unmarked paraphrase that shifts stance or intensity.
• Neutralized register: distinctive voice features smoothed into generic phrasing.

5.9.1.4 Semantic & Interpretive Indicators

• Semantic drift
• Neutralization of emotional tone: transformation of emotions into neutral reporting.
• Over-generalization: replacement of specific cultural markers with universal terms.
• Premature disambiguation/over-specificity: unwarranted resolution of ambiguity where the source permits multiple interpretations.
• Canonical pull: convergence toward widely circulated formulations or precedent versions rather than the given source.
• Domain transfer: insertion of terminology from similar but non-identical contexts.
• Back-translation mismatch: apparent source fidelity that, upon back-translation, reveals lexical or collocational drift not licensed by the source (heuristic red flag, not proof).

The following table lists the qualitative indicators of flattening.
These are cues you can notice just by reading side-by-side: something missing, something dulled, or something added that shifts the voice.
Use this as a checklist: scan the source and mediated text, tick off any indicators you spot, and then note which flattening form they belong to.

Table 2. Qualitative indicators of flattening, grouped by form with description.

5.9.2 Quantitative indicators

The quantitative layer adds efficiency and consistency to the detection process.

• Quantitative indicators MAY be used as diagnostic flags to identify where changes have occurred; they signal areas requiring qualitative investigation, not evidence of flattening itself.
• Quantitative indicators MAY be used as proxies for detection and triage; they are diagnostic aids, not conclusive evidence.
• Where feasible, analysts SHOULD report both a qualitative rationale and any quantitative proxy values used.

5.9.2.1 Cadence indicators:

• Homogenized cadence: measurable as reduced standard deviation of sentence length across aligned segments.
• Flattened pacing: measurable as reduced variance in paragraph length or punctuation frequency.
• Dialogic cadence flattening: measurable as reduced variance in turn length or decreased alternation frequency across aligned conversational turns.

5.9.2.2 Stylistic indicators:

• Loss of idioms: detectable via lexicon match (idiom dictionaries / cultural phrase databases).
• Neutralized register: measurable as drop in type–token ratio (lexical richness).

5.9.2.3 Semantic & interpretive indicators:

• Over-generalization: measurable as increased frequency of high-frequency/general-purpose tokens (e.g., “thing,” “person,” “place”).
• Premature disambiguation: measurable as reduced entropy in word-sense distributions (proxy from embeddings).

The next table lists the quantitative indicators of flattening. These are the same forms, but expressed as patterns you can count: shorter sentences, fewer idioms, less varied vocabulary.
Use this table if you want to back up your qualitative observations with numbers: helpful for reports, but not required for a first pass.

Table 3. Quantitative indicators of flattening, grouped by form with description. Quantitative indicators flag potential areas of change; they do not constitute evidence of flattening without qualitative analysis.

5.9.3 Diagnostic vs. Evidential Use

NFAF employs a two-stage process:

1. Quantitative flags: Identify segments where measurable changes occurred
2. Qualitative analysis: Determine whether flagged changes constitute domain-relevant flattening

Quantitative indicators serve as a triage system, not as proof. A drop in sentence length variance flags potential cadence changes; qualitative analysis determines whether those changes flatten narrative rhythm in that domain context.

The flattening indicators can be read alongside translation “loss markers” (Berman 1985), journalistic “framing shifts” (Hall 1980), and archival debates about authenticity. NFAF formalizes these signals as a domain-crossing taxonomy for AI mediation, ensuring that flattening cannot hide behind fluency. The framework detects and categorizes such changes consistently. Their significance, however, depends on context, goals, and domain expertise: a literary scholar may value preserving sentence-level rhythm variance, while a legal translator may appropriately flatten it for clarity.

5.10 Limitations and Cultural Scope

NFAF categories are grounded in Western narratology, translation studies, journalism, and legal traditions, and therefore carry the imprint of those intellectual lineages. This makes the framework diagnostic rather than universal: what counts as flattening is itself culturally contingent.

Future applications should adapt or extend the rubric with input from non-Western, multilingual, and Indigenous perspectives, testing whether the same dimensions apply, require modification, or misfire entirely. NFAF v0.5 should thus be read as provisional: a shared baseline that enables comparison, but one that must be revised through cross-cultural validation and stakeholder critique.

6. Stages of Analysis

Input Selection --> Transformation --> Alignment --> Drift Assessment & Rubric Scoring at mediation step --> Pattern Analysis --> Human-in-the-Loop Review --> Reporting

The following stages define the normative workflow for applying NFAF. Analysts MUST implement Stages 1–5 and Stage 7, adapting them as needed for their domain. Stage 6 is REQUIRED only in high-stakes contexts (e.g., legal, medical).

• Stage 1 anchors the process through scope declaration, baseline identification, and analysis-level setup.
• Stages 2–6 execute the analytic workflow: generating transformations, aligning source and output, assessing drift, identifying recurring patterns, and, when required, incorporating human review.
• Stage 7 produces the reporting layer, ensuring transparency and accountability.

Together, these stages ensure that flattening is not only detected but also documented with reproducible evidence. Examples in §9 illustrate how these stages may be adapted across literature, journalism, and law.

6.1 Stage 1: Input Selection & Baseline Identification

Analysts MUST implement this stage, though methods (e.g., baseline choice, coding guidelines) may vary by domain.

• Define detailed coding guidelines and decide whether at least two analysts will independently code a subset of materials.
• Defined desired versioning scheme.
• Scope declaration. Analysts MUST declare the corpus scope (languages, domains, time period, model versions if known).
• Source text identification. Establish the source narratives to be transformed.
• Human baseline identification. When available, specify human translations or rewritings for comparison; these baselines SHOULD be included in later stages (§6.3).
• Transformation definition. Define the intended transformation (translation, summarization, rewriting).
• Purpose documentation. Analysts MUST record the purpose and scope of the transformation to contextualize findings.
• Define analysis levels for alignment, drift assessment and rubric scoring.

6.2 Stage 2: Transformation

Analysts MUST document the mediation type (translation, summary, rewrite, etc.). Adaptation to domain conventions is permitted.

• Preprocess source narratives.
• Apply automated or AI-mediated processes to generate alternate versions.
• Preprocessing choices MUST be documented, as they can introduce flattening effects (§5.9).

6.3 Stage 3: Alignment

Analysts MUST perform alignment at a level appropriate to the domain (sentence, paragraph, event, etc.). Confidence levels MUST be declared.

• Map source to output segments** (§5.7 and §5.8).
• Analysts MUST record unmatched spans (omissions), inflation (one-to-many), and fused mappings (many-to-one).
• Granularity defaults: sentence-level (prose), clause-level (legal), turn-level (dialogue). Deviations MUST be justified.
• Alignment manifests (IDs, offsets) SHOULD be released with examples to enable independent verification.
• Human Baseline Comparison: When available, analysts SHOULD include professional human translations of the same source texts. Pilot validation demonstrates systematic divergence between human preservation patterns (positive scores) and AI flattening patterns (negative scores), providing crucial context for interpreting observed changes.

6.3.1 Alignment Confidence Levels

NFAF is diagnostic, not deterministic, analysts MUST declare alignment confidence and document limitations:

• Direct alignment: Clear 1:1 correspondence (sentence/clause level)
• Interpretive alignment: Requires analytical judgment due to restructuring
• Partial alignment: Some segments unmappable due to transformation complexity
• Failed alignment: Structural reorganization prevents reliable correspondence

NFAF findings SHOULD be interpreted within declared confidence limits. Lower confidence does not invalidate analysis but constrains claims about undetected flattening effects.

6.4 Stage 4: Drift Assessment & Rubric Scoring at each mediation step

Analysts MUST assess structural and semantic drift at each mediation step. Scoring MAY be adapted to domain-relevant cues.

• Assess Structural Drift (§5.6.1) and Semantic Drift (§5.6.2) at every mediation step. In multi-step AI mediation processes, flattening effects may compound; therefore, drift and rubric scoring should be performed at each stage, not only on the final output.
• Score and record each dimension at segment level.
• Analysts MUST examine shifts in dialogue, omissions, event collapsing, tone, and cultural/emotional signals.

Quantitative Indicators --> Flag Changes --> Qualitative Analysis --> Domain Assessment --> Flattening Diagnosis

6.5 Stage 5: Pattern Analysis

Analysts MUST analyze recurring flattening patterns. Multiple-analyst agreement SHOULD be measured where applicable.

• If multiple analysts, measure inter-annotator agreement using standard metrics (Cohen's kappa, etc.); though, perfect agreement is not expected.
• Record recurring flattening patterns across transformations.
• Record recurring flattening patterns across transformations.
• Analysts SHOULD contrast NFAF findings with standard NLP metrics (BLEU, ROUGE, perplexity) to highlight narrative-level effects not captured by surface metrics.

6.6 Stage 6: Human-in-the-Loop Review (optional except in high-stakes contexts)

Optional in general practice, but MUST be included for high-stakes contexts (e.g., law, medicine, archival preservation).

• Analysts MAY validate findings through targeted human evaluation.
• For high-stakes contexts (e.g., legal, medical), human review MUST be included.

6.7 Stage 7: Reporting

Analysts MUST produce a reporting layer documenting evidence, rationale, and limitations. Format may be adapted by domain (tables, narrative notes, visualizations).

• Summarize observed patterns and contributing factors.
• When omissions occur, outputs MUST be labeled as summaries or rewrites and accompanied by an alignment manifest (§5.7).
• Omission without disclosure MUST be treated as flattening and constitutes a material misrepresentation in the interpretive/ethical sense.

The next table lays out the NFAF stages of analysis.
These stages show the order of work: how to prepare your input, run the mediation, align the texts, identify drift, and finally report your findings.
Use this table as a roadmap: follow the stages in sequence, and note that Stage 6 (Human-in-the-Loop Review) is optional in ordinary cases but mandatory in high-stakes contexts.

Table 4. Stages of analysis in the Narrative Flattening Analysis Framework (NFAF).

7. Rubric

NFAF rubric dimensions detect systematic changes but do not determine acceptability thresholds. Domain experts must interpret whether observed changes constitute problematic flattening based on context-specific values and requirements. Metrics defined here are diagnostic indicators validated against human translation baselines, and MUST always be paired with qualitative evidence. They are not prescriptive thresholds.

The dimensions below should be interpreted in relation to the recurrent forms of flattening defined in §5.3 (erosion, inflation, inference, omission/compression). Together, forms and dimensions provide a layered framework: forms describe how flattening manifests; dimensions describe where it is observed. If AI mediation normalizes inputs through a Centroid Pivot (§5.4.2), systemic signatures are commonly observed across the following dimensions:

• Cadence: variance drops (e.g., English-like pacing).
• Lexical/Stylistic: substitution of culture-specific items with globally fluent paraphrase.
• Cultural Register: dilution of local or culturally marked terms.
• Semantic/Interpretive: shifts toward Anglophone priors (e.g., canonical pull, premature disambiguation).

Analysts MUST disclose observed shifts, and reports MUST pair quantitative Δ values with qualitative examples (e.g., idiom paraphrase, loss of irony).

The rubric dimensions define the core lenses for analyzing flattening.
Each dimension corresponds to a specific type of narrative signal: rhythm, vocabulary richness, figurative language, cultural references, affect, or structural coverage.
The table shows (a) the signals to check, (b) guidance for interpretation, and (c) reporting obligations to ensure analytical rigor and ethical responsibility.
Use this as your anchor: every observed indicator can be tied back to one or more rubric dimensions.

Table 5. Rubric dimensions for analyzing narrative flattening, with signals, guidance, and reporting responsibilities.
Δ convention: Output − Source (positive = more in output; negative = less). Example: if sentence-length variance drops from 42.1 → 18.6, Δσ² = −23.5, a flattening cue.

7.1 NFAF Rubric: Cadence & Rhythm

Signals:

• Mean and variance of sentence length.
• Punctuation cadence (commas, em-dashes, ellipses per 1k tokens).
• Figurative or imagery clusters that shape narrative flow.

Testable Predictions:

1. Variance reduction. Sentence length variance (σ²) commonly decreases after AI mediation, converging toward English-like pacing. Analysts SHOULD compare against source-genre baselines before concluding flattening.
2. Prosodic flattening. Distinctive rhythm markers (em-dashes, ellipses, enjambment) are replaced by default punctuation (commas, periods). Analysts MUST disclose such shifts as cadence loss.
3. Peak loss. Figurative “bursts” (clusters of metaphor, irony, lyricism) appear less frequently or are redistributed more evenly across the text. Analysts SHOULD document the redistribution as flattening of intensity.

Normative Directive:
Flattening of cadence MUST be disclosed as a loss of narrative voice. Treating rhythmic simplification as “clarification” constitutes flattening and erases meaning.

Anchors:

• Stylistics: cadence as a carrier of meaning, not ornament (Leech & Short 1981).
• Authorship studies: variance as a stylistic fingerprint (Burrows 2002; Stamatatos 2009).
• Translation studies: critique of “rationalization” (Berman 1985).
• NLP: systemic convergence in embeddings (Hamilton et al. 2016); pivot layers in NMT (Johnson et al. 2017).

7.2 NFAF Rubric: Lexical Richness

Signals:

• Type–Token Ratio (TTR).
• Measure of Textual Lexical Diversity (MTLD).
• Hapax legomena share (unique once-used words).
• Rare-word rate (frequency quantile relative to source or a reference corpus).

Testable Predictions:

1. Diversity reduction. TTR and MTLD scores commonly decrease in AI-mediated outputs, showing convergence on mid-frequency vocabulary. Analysts SHOULD test against source-genre baselines.
2. Rare-word suppression. Hapax legomena and culturally marked lexemes decline, substituted with high-frequency synonyms. Analysts MUST document such substitutions when they alter cultural or stylistic signals.
3. Register neutralization. Distinctive technical, regional, or sociolectal vocabulary is replaced by globally accessible terms. Analysts SHOULD disclose such neutralization.
4. Overuse of “safe” lexemes. A limited set of frequent words (e.g., “important,” “significant,” “good”) appears at higher rates than in the source. Analysts SHOULD record elevated repetition as a flattening cue.

Normative Directive:
Suppression or substitution of rare and culturally marked lexemes MUST be treated as erasure. Ethical reporting requires making these substitutions visible. Analysts SHOULD interpret lexical diversity within the typological context of the language (inflectional vs. analytic).

Anchors:

• Stylistics: vocabulary as a stylistic marker (Leech & Short 1981).
• Translation studies: translational norms favoring common lexical items (Toury 1995).
• Computational stylistics: distinctive lexical distributions for authorship attribution (Burrows 2002).
• NLP: centroid bias privileging high-frequency English lexicon (Bender et al. 2021).
• Corpus linguistics: measuring lexical diversity across corpora (McEnery & Hardie 2012).

7.3 NFAF Rubric: Figurative Density

Signals:

• Idiom frequency per 1k tokens.
• Simile markers (e.g., “like/as a,” “como un/una,” “comme un/une”).
• Metaphor candidates (rule-based or embedding-based detection if available).
• Clusters of figurative language that drive imagery, irony, or tone.

Testable Predictions:

1. Idiom loss. Culture-specific idioms are commonly paraphrased into generic phrases (e.g., tirar la toalla → “to give up”). Analysts MUST disclose such substitutions.
2. Simile simplification. Similes with culturally rich referents are replaced with neutral or universal comparisons. Analysts SHOULD document the simplification as flattening.
3. Metaphor reduction. Extended metaphoric structures (multi-clause conceits, allegories) are shortened or omitted. Analysts MUST record these reductions where they alter figurative signal.
4. Density smoothing. Peaks of figurative intensity (e.g., lyric passages in literature, rhetorical flourishes in journalism or law) are redistributed into more even, literal prose. Analysts SHOULD note redistribution patterns as flattening cues.

Normative Directive:
Paraphrasing idioms, similes, or metaphors into plain text MUST be treated as a loss of figurative signal. Ethical reporting requires documenting these shifts. Analysts SHOULD supplement automated tools with manual annotation, since existing systems under-detect culturally marked figuratives.

Anchors:

• Translation studies: figurative “clarification” as flattening (Berman 1985).
• Stylistics: metaphor and idiom as core stylistic markers (Leech & Short 1981).
• Digital Humanities: figurative density as quantifiable signal in large corpora (Jockers 2013; Moretti 2013).
• NLP: AI bias toward literal paraphrase when figurative usage is uncertain (Shutova 2010; Dankers et al. 2019).

7.4 NFAF Rubric: Cultural Register

Signals:

• Named entities tied to locale (places, foods, institutions, idioms).
• Culture-specific lexicon and sayings.
• Percentage of culture-specific terms replaced by generic, universal, or Anglophone-oriented terms.
• Presence or absence of honorifics, politeness markers, or culturally specific discourse cues.

Testable Predictions:

1. Entity dilution. Local place names, foods, or institutions are generalized (e.g., bodega → “corner store,” soba → “noodles”). Analysts MUST disclose these substitutions.
2. Discourse marker shifts. Loss or overuse of culturally distinctive connectives (e.g., pues in Spanish, en fait in French). Analysts SHOULD document these shifts.
3. Idiom neutralization. Proverbs and culturally distinctive sayings are paraphrased into plain or global equivalents. Analysts MUST treat these as flattening.
4. Honorific loss. Politeness markers (e.g., usted, -san, monsieur) are omitted or flattened into direct address. Analysts MUST flag these omissions.
5. Centroid language substitution. Target text exhibits features from the dominant centroid language: for example, Anglophone metaphors, collocations, or formatting (e.g., U.S.-style date/time, “Mr.” for non-English honorifics). Analysts MUST report such substitutions as cultural overwriting.
6. Centroid Pivot Imprint. Even in non-English ↔ non-English transformations, outputs may carry stylistic defaults from the pivot language. This imprint MAY vary by system and dataset. At present, it often manifests as English-like discourse markers (e.g., en realidad, como resultado in Spanish; overuse of en fait in French). Analysts SHOULD treat these as indicative of systemic pivot influence rather than source/target features.

Normative Directive:
Shifts in discourse markers, honorifics, or culturally specific lexicon MUST be flagged. Such substitutions are not neutral; they constitute flattening and risk overwriting situated cultural identity with centroid defaults.

Anchors:

• Translation studies: ethnocentric reduction of idioms and references (Berman 1985); translational norms under systemic pressure (Toury 1995); domestication as erasure of local markers (Venuti 1995).
• Globalization: flattening of small/local languages through lingua franca mediation (Cronin 2003).
• NLP/MT: systemic English dominance in multilingual models (Johnson et al. 2017; Bender et al. 2021).

7.5 NFAF Rubric: Emotional Intensity

Signals:

• Evaluative/affect lexicon counts (positive/negative adjectives, intensifiers).
• Sentiment polarity and subjectivity measures.
• Variability of affect (standard deviation of sentiment scores across segments).
• Use of exclamatives, modal verbs, rhetorical questions as emotional carriers.
• Sarcasm/irony markers (contrastive cues, hyperbole, scare quotes) that convey affect indirectly.

Testable Predictions:

1. Neutralization of affect. Strongly positive or negative terms are commonly replaced by moderate descriptors (e.g., horrifying → “bad,” exquisite → “beautiful”). Analysts MUST disclose such substitutions.
2. Range compression. Standard deviation of affect scores often decreases; emotional highs and lows flatten into mid-intensity prose. Analysts SHOULD compare against source-genre baselines.
3. Removal of markers. Exclamatives, modal intensifiers (must, should), or rhetorical questions are omitted or converted to declaratives. Analysts MUST flag these as flattening.
4. Cultural calibration loss. Affect signals are normalized without regard to cultural conventions (e.g., ¡Hola! vs. Hello! vs. ！). Analysts SHOULD document when mediation imposes a uniform mid-level register.
5. Tone smoothing across genres. Journalism summaries downplay outrage or urgency; literary passages lose lyric peaks of sorrow/joy; legal rhetoric loses emphatic certainty. Analysts SHOULD record such genre-dependent smoothing.

Normative Directive:
Neutralization or compression of affect MUST be treated as flattening of narrative arcs. Emotional resonance cannot be erased invisibly; analysts MUST disclose these changes. Analysts SHOULD calibrate affect tools against source-language baselines or use manual annotation where automated tools are Anglophone-calibrated.

Anchors:

• Narratology: mood and tone as narrative discourse dimensions (Genette 1980).
• Translation studies: tonal reduction via “ennoblement” or “clarification” (Berman 1985); domestication into neutral register (Venuti 1995).
• Digital Humanities: sentiment arc measurement (Ramsay 2011; Jockers 2013).
• NLP: affect lexicons (Pennebaker et al. 2001; Fast et al. 2016); moderation toward neutrality in LLM sentiment outputs (recent work).

7.6 NFAF Rubric: Coverage & Structure

Signals:

• Segment retention ratio (kept/total).
• Compression ratio (tokens_out ÷ tokens_in).
• Count of unmatched spans (omitted, fused, or inflated segments).
• Preservation of discourse relations (contrast, cause, elaboration via Rhetorical Structure Theory (RST) or Universal Dependencies (UD) markers).
• Integrity of quotations, speaker turns, and attributions.

Testable Predictions:

1. Segment loss. AI-mediated outputs commonly omit or fuse multiple source segments, reducing overall coverage. Analysts MUST flag such omissions.
2. Compression bias. Outputs exhibit higher compression ratios than human-mediated equivalents, especially in summarization. Analysts SHOULD compare across baselines.
3. Discourse link collapse. Explicit markers of causality, contrast, or elaboration are dropped (e.g., sin embargo, parce que, “however”). Analysts MUST disclose such losses.
4. Dialogue flattening. Multi-turn exchanges are collapsed into single-speaker paraphrase. Analysts MUST report such flattening.
5. Structural drift. Narrative scaffold (event sequencing, clause nesting, paragraphing) is altered without disclosure, producing fluent but structurally divergent outputs. Analysts MUST flag undisclosed drift.
6. Multimodal omission. When inputs include images, graphs, or visuals, mediated outputs may omit them. Analysts SHOULD record such omissions explicitly.
7. Longitudinal drift. Chronologies are compressed or events reordered, altering historical or archival integrity. Analysts MUST report such shifts.
8. Integrity of formatting/paratext. Footnotes, references, or headings (e.g., legal citations) are dropped. Analysts MUST treat such omissions as structural flattening.

Normative Directive:
Omissions, re-orderings, or structural drift without disclosure MUST be treated as flattening. Such changes mislead readers about fidelity and compromise accountability.

Anchors:

• Discourse theory: discourse relations as narrative scaffold (Mann & Thompson 1988; van Dijk 2008).
• Translation studies: omission and rationalization as systemic tendencies (Berman 1985).
• Journalism studies: framing and omission reshape stance (Hall 1980).
• NLP: compression vs. fidelity trade-offs in summarization (Hamilton et al. 2016; ROUGE, BLEU).
• Archival science: provenance and integrity principles highlight why omissions compromise accountability.

7.7 NFAF Rubric: Rubric Extensions and Domain-Specific Quantification

The NFAF rubric is defined across six dimensions: cadence & rhythm, lexical richness, figurative density, cultural register, emotional intensity, and coverage & structure. These indicators are deliberately domain-agnostic, allowing them to be observed across literary texts, journalistic accounts, legal registers, or multimodal narratives.

While the framework itself remains fixed, the rubric admits domain-specific extensions in the form of quantification schemes. Such extensions transform qualitative indicators into reproducible scoring systems tailored to a research context. Examples include:

• Journalism and News ratios of global newswire phrasing versus local idiomatic reporting to quantify stance homogenization.
• Translation Studies a signed scale per dimension (e.g., −5 … +5) to capture subtractive versus additive drift in AI-mediated translation.
• Historical and Archival Texts coverage scored by omissions or condensations in AI-generated summaries of primary sources.
• Legal Texts lexical scoring via shifts in statutory anchors versus paraphrase density.
• Multimodal Narratives (Images/Film) cadence measured as variance in frame or scene sequencing; coverage as motif preservation.

These extensions do not alter the framework itself; they are layered on top of the existing rubric to make flattening measurable in a given domain. NFAF thus remains a general diagnostic framework, while individual projects demonstrate how its dimensions can be operationalized through quantification.

Preliminary validation demonstrates the rubric’s diagnostic capability. Across pilot corpora, AI translations show systematic negative scores on figurative density, cadence, and cultural register dimensions, while human translations show positive scores, indicating preservation or enhancement of source features. This empirical distinction validates NFAF’s theoretical claims about systematic AI convergence effects and motivates the foundations detailed in §8.

8. NFAF Foundations

NFAF situates itself within and extends prior debates in narratology, translation studies, stylistics, and computational linguistics. Some categories are adapted from existing traditions (e.g., Drift from linguistics, Alignment from corpus studies); others are defined within this framework (e.g., Narrative Flattening, Centroid Pivot). This distinction is deliberate: - Adapts = terms reframed for AI mediation but grounded in prior scholarship. - Defines = terms coined in NFAF to capture effects not previously formalized.

Preliminary validation (§7.7) confirms that these constructs are not only theoretical but empirically observable: AI translations show systematic negative scores across figurative density, cadence, and cultural register dimensions, while human translations show positive scores, indicating preservation or enhancement of source features. This empirical distinction validates the framework’s theoretical claims about systematic AI convergence effects.

• Centroid Pivot and Precedents.
  NFAF defines Centroid Pivot (CP) as a general mechanism of normalization toward dominant defaults. It is a normative construct for analysis, not a claim about specific system internals. Precedents include:

  • Relay translation (Toury 1995; Santoyo 2009), where texts pass through intermediate languages.
  • Interlingua / pivot layers in machine translation (Koehn 2009), where systems map to intermediate representations.
  • Globalization pivots in translation flows (Cronin 2003), where certain languages act as systemic hubs.

  NFAF extends these by articulating CP as:

  1. A mechanism not limited to translation but spanning modalities (text, image, sound).
  2. A process that may be internal (inside a model or cognition) or external (relay between mediators).
  3. A force that is not language-bound: in medieval Spain, Castilian Spanish served as a pivot; in contemporary AI, the statistical default is often English; in future systems, Mandarin or Hindi could dominate.

• Narrative Flattening and Cultural Heritage.
  NFAF defines Narrative Flattening as systemic reduction or distortion of distinctiveness in voice, rhythm, cultural register, or emotional intensity during mediation. While translation studies (Venuti 1995; Berman 1985) describe analogous losses under domestication or the “negative analytic,” NFAF specifies flattening as a reproducible, diagnosable phenomenon across mediation types, including AI rewriting and summarization.

• Narratology (Genette 1980; Bal 1985; Ryan 2004).
  Established the study of discourse, framing, and sequencing as carriers of meaning.

• Translation Studies (Venuti 1995; Berman 1985; Toury 1995).
  Foregrounded ethics of mediation, domestication, and invisibility. NFAF acknowledges these debates but emphasizes how AI mediation amplifies them through scale and recursive feedback loops.

• Digital Humanities (Drucker 2014; Ramsay 2011; Jockers 2013).
  Advanced computational and interpretive reading, critiquing assumptions of neutrality.

• Stylistics & Authorship Studies (Leech & Short 1981; Burrows 2002; Stamatatos 2009).
  Quantified style through cadence, vocabulary, and authorship markers.

• Media & Culture (Hall 1980).
  Framing theory and encoding/decoding show how mediation shapes stance.

• AI / NLP (Papineni et al. 2002; Lin 2004; Koehn 2009; Hamilton et al. 2016; Mohammad & Turney 2013; Bender et al. 2021; Mitchell 2021).
  Exposed bias and convergence in large models; introduced evaluation standards (BLEU, ROUGE); developed sentiment lexicons. NFAF complements, rather than replaces, these surface metrics by focusing on narrative voice, cultural register, and emotional resonance.

• Mechanism & Mediation (Cronin 2003; Johnson et al. 2017).
  Identified globalization pressures and pivot mechanisms that flatten linguistic and cultural signals.

• Alignment Studies (Gale & Church 1993).
  Introduced statistical methods for aligning bilingual corpora, foundations for modern parallel analysis.

The following table situates the NFAF rubric dimensions within established disciplinary traditions.
Each field provides conceptual anchors (narratology, translation studies, stylistics, digital humanities, and more) that inform how flattening is understood and measured.
Use this table as a map: it shows that the rubric dimensions are not arbitrary but extend and recombine existing debates across theory, method, and practice.

Table 6. Disciplinary anchors and their contributions to NFAF rubric dimensions.

Together, these anchors demonstrate that NFAF is not limited to any single discipline: it translates long-standing humanistic concerns into a replicable diagnostic for the AI-mediated era. Preliminary validation confirms this: AI translations show systematic negative scores across figurative density, cadence, and cultural register dimensions, while human translations show positive scores, indicating preservation or enhancement of source features. This empirical distinction validates the framework’s theoretical claims about systematic AI convergence effects.

9. Applications: NFAF AI-mediated Flattening - Illustrative, not exhaustive.

In the following examples, NFAF conventions are applied to demonstrate how flattening can be identified in practice.
For clarity, each qualitative alignment table is followed by an Observed block (what is visible in the text), and each quantitative metrics table is followed by an Interpretation block (how to read the numbers as diagnostic cues).
These labels are conventions for the examples only; they are not part of the core NFAF specification.
The aim is to help the reader see how qualitative and quantitative evidence complement one another in detecting narrative flattening.

NFAF does not issue verdicts. It equips stakeholders with categories, indicators, and rubrics to make transformation effects visible. The diagnosis of flattening is always situated and interpretive: stakeholders decide, based on evidence, whether observed changes constitute flattening in a given domain or mediation type.

Model Information

Model information is at ./examples_ejemplos/model_information.txt

• Model A: ChatGPT
• Model B: Claude
• Model C: Gemini
• Model D: DeepL

Generic NFAF Workflow (Normative)

• Stage 1: Input Selection & Baseline Identification
• Stage 2: Transformation (AI-mediated)
• Stage 3: Alignment (level adapted to domain)
• Stage 4: Drift Assessment & Rubric Scoring (structural + semantic)
• Stage 5: Pattern Analysis (recurring flattening)
• Stage 6: Human-in-the-Loop Review (optional, REQUIRED in high-stakes)
• Stage 7: Reporting (domain-adapted outputs)

9.1 Literary Rewrite (Austen, Emma)

NFAF evaluates outputs, not actors; the aim is transparency around transformation effects.

• Domain: Literary
• Focus: Stylistic flattening (voice/cadence)
• EX01 version: v01 Source: Jane Austen, Emma (1815), Chapter 1 excerpt (public domain).
  Mediation: Modernized rewrite by three models.
  Data: Data files are located at ./examples_ejemplos/EX01_Literature_Literatura

Summary Finding:
Rewrites produced fluent modern English but erased Austen’s irony and narrative rhythm. Omissions were not disclosed, constituting flattening of Coverage & Structure (§7.6).

Application Glossary (for the Emma ch1 Paragraph 5 case study):

• Rewrite: 1–1 mapping.
• Summary: Omission with disclosure.
• Missing: Null alignment.
• Inflated rewrite: Expansion with padding.

Sequence

• Input Selection & Baseline Identification: Emma ch1 (paragraph-level baseline)
• Transformation: AI-mediated rewrite
• Alignment: paragraph-level
• Drift Assessment & Rubric Scoring: quantitative + qualitative
• Pattern Analysis: recurring stylistic flattening
• Human-in-the-Loop Review: Optional. Alignment and analysis performed by human analyst.
• Reporting: tables + narrative notes

Alignment Excerpt (Emma ch1 Paragraph 5)

The next table shows a worked alignment example using Austen’s Emma (chapter 1, paragraph 5).
The source passage is placed side by side with three AI-mediated rewrites.
Use this table to see how flattening is detected in practice: align source to output, note shifts in cadence, idioms, or omissions, and then trace those observations back to the rubric dimensions introduced earlier.

Table 7. Alignment excerpt from Austen’s Emma, showing source against three AI rewrites.

Observed:

• Model A (Omission/Compression) compresses to a hybrid: half rewrite, half summary (drops dining/nap scene into generalized reflection).
• Model B (Erosion - Cadence)preserves all narrative beats (true rewrite, but cadence is flattened).
• Model C (Inflation) preserves sequence but inflates with over-specific phrasing (“continuous mournful thought”).

Cadence Metric (Emma ch1 Paragraph 5)

This table demonstrates how NFAF can make cadence shifts visible with simple metrics.
It compares mean sentence length and variance between the source and model outputs.
The numbers are not statistical proof: they illustrate how smoothing cadence (lower variance) signals flattening of rhythm.
Use this as a diagnostic aid: if variance collapses, the text has lost its ebb and flow.

Note: This example isn't meant to be statistical proof - it's a demonstration of how the framework operates.

*Table 8. Cadence metrics for Austen’s Emma, showing reduced variance and flattened rhythm in model rewrites.

Interpretation: Sentence-length variance drops sharply in all rewrites, demonstrating homogenized Cadence & Rhythm (§7.1). Austen’s rhythmic “peaks and valleys” are flattened into more even prose. Qualitative evidence shows loss of irony and tonal shifts, linking also to Emotional Intensity (§7.5).

Variance reduction is a diagnostic cue, not proof in isolation. Stakeholders using NFAF must combine quantitative measures with qualitative evidence to determine whether flattening has occurred.

9.2 Journalism & News (Summarization)

NFAF evaluates outputs, not actors; the aim is transparency around transformation effects.

In contrast to the literary rewrite example in §9.1, where flattening appeared as the loss of cadence, irony, and stylistic rhythm, journalism demonstrates a different pattern. Here, the mediation is summarization, and the domain is news reporting across regions. Flattening does not emerge as stylistic erosion but as the neutralization of editorial stance, cultural references, and actor voices, resulting in a homogenized “wire-style” narrative. This illustrates how the same rubric dimensions apply across domains, even though the signals of flattening differ.

• Domain: Journalism & News
• Focus: Cultural flattening (journalistic stance)
• EX02 version: v01

Source: Six outlets (Guardian, Fox, El País, BBC, AP, Reuters) reported on the same U.S. court ruling, each with its own cultural stance and editorial register.
Mediation: Summarization by one model.
Data: Data files are located at ./examples_ejemplos/EX02_Journalism_Noticias

Summary Finding:
Claude’s summaries consistently neutralized outlet-specific differences, compressing coverage (though less severely than other systems), thinning rather than eliminating quotes, paraphrasing evaluative language, and erasing or generalizing cultural markers (“migrantes,” “Everglades,” “controversial”). In one case, stance markers were even added (“Trump-appointed judge”), showing a mild canonical pull. This demonstrates cultural flattening: the systematic reduction of diverse journalistic voices into a single, homogenized register.

Application Glossary (for the Journalism & News case study):

• Stance markers: Partisan or evaluative labels (e.g., Obama-appointed judge, open-borders activists).
• Cultural register: Localized or regional references (e.g., Everglades, Miccosukee Tribe, migrantes).
• Editorial diversity: Distinct journalistic voices shaped by outlet, region, or political orientation.
• Centroid news register: Neutral, standardized “wire” style prose resembling AP/BBC tone.

Sequence

• Input Selection & Baseline Identification: Guardian, Fox, El País, BBC, AP, Reuters
• Transformation: AI-mediated summarization
• Alignment: article-level
• Drift Assessment & Rubric Scoring: quantitative + qualitative
• Pattern Analysis: cultural flattening patterns
• Human-in-the-Loop Review: Optional. Alignment and analysis performed by human analyst.
• Reporting: tables + narrative notes

Alignment Excerpts (Six outlets - Journalism & News (Summarization))

The following table shows alignment excerpts for six outlets reporting on the same U.S. court ruling.
Each outlet presents its coverage with a distinct editorial register and cultural stance.
The Model B summarization, shown alongside, demonstrates how diverse journalistic voices are compressed into a uniform style.

Table 9. Alignment excerpts (six outlets) with one AI summarization.

Rubric Analysis ( Six outlets - Summary Journalism & News)

The next table applies the NFAF rubric dimensions to the journalism case.
It highlights how source signals (regional stance, figurative density, cultural references, emotional tone, and structural coverage) are transformed in summarization.

Table 10. Rubric Analysis for Journalism & News example using six outlets by one AI summarization.

This table provides outlet-specific notes that trace exactly what was lost or neutralized in each case.
By comparing sources to Model B outputs, the table makes visible the omission of partisan framing, cultural markers, and regional context.
Together, these notes show how plural journalistic registers converged on a single, neutralized narrative.

Table 11. Flattening Notes for Journalism & News (Summarization) example using six outlets.

Observed:

• Model B (Claude) neutralizes stance (Erosion), compresses coverage (Omission/Compression), and erases cultural specificity (Cultural Register flattening).
• Distinct voices: progressive (Guardian), conservative (Fox), European (Reuters), Spanish/Latino (El País), and neutral/procedural (BBC, AP) collapse into a single centroid news register.
• While Claude occasionally retains one or two direct quotes, the overall pattern tends toward neutralization, partisan identifiers omitted, and cultural references erased.
• In journalistic contexts, flattening manifests not as loss of literary style (as in Emma ch1) but as the erasure of editorial diversity and stance plurality in the public sphere.
• This aligns with the rubric dimensions of Cultural Register (§7.4), Emotional Intensity (§7.5), and Coverage & Structure (§7.6), alongside cadence convergence (Cadence & Rhythm §7.1).

NFAF does not automatically classify shorter length as flattening. Instead, stakeholders may diagnose flattening when compression disproportionately erases stance markers, cultural references, or actor voices essential to narrative diversity.

Note: Summarization inherently involves compression. NFAF does not prescribe thresholds for acceptable loss; responsibility for diagnosis resides with the stakeholder.

9.3 Law (Case T-612/17 para. 27 & 704 - Translation)

NFAF evaluates outputs, not actors; the aim is transparency around transformation effects.

• Domain: Law
• Focus: Jurisdictional flattening (legal register + statutory anchors)
• EX03 version: v01

Source: Court of Justice of the EU, Case T-612/17 (Google v Commission), Judgment excerpt, para. 27 and 704 (public domain).
Mediation: EN → ES by DeepL, with comparison to official EUR-Lex ES translation, then back-translation (ES → EN by DeepL). Data: Data files are located at ./examples_ejemplos/EX03_Law_Legal
Summary Finding:
In this example, translation drift emerges less through omission and more through tense/aspect smoothing and register dilution. In para. 27, DeepL’s choice of llegaba (imperfect) instead of the legal pluperfect había llegado weakens the finality of the Commission’s provisional conclusion. In para. 704, the official Spanish preserves civil law procedural anchors (petitum, subsidiario), while DeepL collapses these into the Anglophone-generic “alternative claim.” Across both, the outputs converge toward a simplified register that can attenuate the juridical weight of decisions. Whether this constitutes flattening is a stakeholder judgment. The stakeholders need to evaluate this instance of jurisdictional flattening.

Sequence

• Input Selection & Baseline Identification: Court of Justice of the EU, Case T-612/17 (Google v Commission)
• Transformation: AI-mediated translation + back-translation
• Alignment: paragraph-level
• Drift Assessment & Rubric Scoring: qualitative (tense, aspect, register)
• Pattern Analysis: jurisdictional flattening
• Human-in-the-Loop Review: REQUIRED in high-stakes (legal context), note - alignment and analysis perform by human
• Reporting: rubric table + flattening notes

This table presents selected alignment excerpts from the European Commission’s antitrust decision in the Google Shopping (Court of Justice of the EU, Case T-612/17) case. The passages were first translated from English into Spanish using DeepL, then back-translated into English using DeepL. Placing the source, translation, and back-translation side by side makes it possible to detect subtle divergences at the level of aspect, tense, and legal register. These small shifts illustrate how even high-fidelity translation systems may reconfigure binding force in legal texts.

Table 12. Alignment excerpt (Law, Para. 27 & 704), showing EN source, official ES, DeepL ES, and back-translation.

Rubric Analysis (Case T-612/17 para. 27 & 704 - Translation)

The following rubric table applies the NFAF categories to the law example, focusing on structural alignment, semantic drift, and register. Unlike literature or journalism, the legal domain foregrounds precision in modality and aspect: a single verb choice may carry significant interpretive weight. The analysis highlights how DeepL’s outputs handle these features, and where flattening may emerge.

Table 13. Rubric Analysis for Law (Translation, para. 27 & 704, EN→ES→EN via DeepL).

Observed:

• Tense/Aspect Shift: Reached (perfective, decisive) → llegaba (imperfective, ongoing). This shift weakens juridical finality by recasting a settled conclusion as a process still in motion.
• Register Drift: Controvertidas (specialized legal register) → en cuestión (everyday descriptive). The specialized judicial tone is diluted into a generic phrasing.
• Connector Flattening: Por ende (formal, judicial connective) → por lo tanto (neutral logical link). A procedural marker is reduced to a common logical transition, flattening institutional register.
• Agency Drift (Back-EN): Source stresses the act of reaching a provisional conclusion; back-translation reframes this as provisionally concluding, shifting the provisionality from the status of the conclusion to the Commission’s action.

This table synthesizes observed divergences into interpretive notes. The focus is not on overt mistranslation but on jurisdictional flattening: a gradual leveling of legal force through subtle choices in tense and register. The notes stress that such flattening does not nullify meaning but may dilute the perception of finality, authority, or closure within the legal discourse. These examples demonstrate how NFAF can be extended beyond literature and journalism to systematically interrogate flattening in the legal domain.

Table 14: Flattening Notes for Law (Translation & Back-Translation)

Interpretation:
In this example for the legal domain, flattening does not appear as stylistic erosion (as in literature) or stance neutralization (as in journalism). Instead, it manifests as jurisdictional flattening: the subtle weakening of legal force through aspectual smoothing, register dilution, and loss of procedural anchors. An imperfect verb (llegaba) or the substitution of petitum with “alternative claim” does not erase meaning, but it shifts the perceived finality and authority of the text toward a more generic Anglophone legal style. Such drift is not inherently “error,” but it carries stakes: legal stakeholders must judge whether the flattening materially alters interpretation or enforceability.

9.4 Cross-Domain Applications

NFAF is not limited to the examples in (§9); it can be applied across different domains, each with its own focus, mediation type, and common flattening pattern.

The following table illustrates representative domains where NFAF has been or can be applied. Each domain suggests a likely mediation type and flattening pattern. These are illustrative only and are not exhaustive or prescriptive.

Table 15. Cross-domain applications of NFAF, showing representative domains, mediation types, and example flattening patterns (illustrative only).

10. Extensions: NFAF Domain-Emergent Flattening Across Domains - Illustrative, not exhaustive.

While NFAF v0.5.0 focuses on AI-mediated flattening, the framework also highlights the potential to study domain-emergent flattening: cases where professional practices themselves flatten narrative voice, stance, or meaning.
These extensions are illustrative only and outside the normative scope of this document. Future versions (e.g., NFAF-J, NFAF-L, NFAF-P) may formalize and add different domains.

10.1 Journalism (NFAF-J) - Illustrative

• Signals: Balance markers, stance framing, partisan vs. neutral lexicon.
• Pattern Statement: Journalistic practices may create false balance, neutralize stance, or converge reporting into a globalized news register.
• Status: Out of scope for NFAF v0.5.0; possible extension.

10.2 Law (NFAF-L) - Illustrative

• Signals: Testimony scaffolding, legal citations, register fidelity.
• Pattern Statement: Judicial practices may compress testimony, omit narrative context, or erode nuance in written opinions.
• Status: Out of scope for NFAF v0.5.0; possible extension.

10.3 Politics (NFAF-P) - Illustrative

• Signals: Issue framing, sloganization, binary simplification.
• Pattern Statement: Political discourse often simplifies complex issues into binary framings (e.g., for/against), flattening deliberative nuance.
• Status: Out of scope for NFAF v0.5.0; possible extension.

11. References

Bal, Mieke. 1985. Narratology: Introduction to the Theory of Narrative. Toronto: University of Toronto Press.

Bender, Emily M., Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’21).

Berman, Antoine. 2000. Translation and the trials of the foreign. In L. Venuti (Ed.), The Translation Studies Reader (pp. 284–297). Routledge.

Burrows, John. 2002. “Delta: A Measure of Stylistic Difference and a Guide to Likely Authorship.” Literary and Linguistic Computing 17 (3): 267–287.

Cronin, Michael. 2003. Translation and Globalization. London: Routledge.

Dankers, Verna, Marek Rei, and Ekaterina Shutova. 2019. “Modelling the Interplay of Metaphor and Emotion through Multitask Learning.” In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), 2218–2229.

Drucker, Johanna. 2014. Graphesis: Visual Forms of Knowledge Production. Cambridge, MA: Harvard University Press.

Fast, Ethan, Binbin Chen, and Michael S. Bernstein. 2016. “Empath: Understanding Topic Signals in Large-Scale Text.” In Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems, 4647–4657. New York: ACM.

Gale, William A., and Kenneth W. Church. 1993. “A Program for Aligning Sentences in Bilingual Corpora.” Computational Linguistics 19 (1): 75–102.

Genette, Gérard. 1980. Narrative Discourse: An Essay in Method. Ithaca: Cornell University Press.

Hall, Stuart. 1980. “Encoding/Decoding.” In Culture, Media, Language. London: Routledge.

Hamilton, William L., Jure Leskovec, and Dan Jurafsky. 2016. “Diachronic Word Embeddings Reveal Statistical Laws of Semantic Change.” In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (ACL).

Johnson, Melvin, Mike Schuster, Quoc V. Le, Maxim Krikun, Yonghui Wu, Zhifeng Chen, Nikhil Thorat, et al. 2017. “Google’s Multilingual Neural Machine Translation System: Enabling Zero-Shot Translation.” Transactions of the Association for Computational Linguistics 5: 339–351.

Jockers, Matthew L. 2013. Macroanalysis: Digital Methods and Literary History. Urbana: University of Illinois Press.

Koehn, Philipp. 2009. Statistical Machine Translation. Cambridge: Cambridge University Press.

Leech, Geoffrey, and Mick Short. 1981. Style in Fiction. London: Longman.

Lin, Chin-Yew. 2004. “ROUGE: A Package for Automatic Evaluation of Summaries.” In Text Summarization Branches Out: Proceedings of the ACL-04 Workshop, 74–81.

Mann, William C., and Sandra A. Thompson. 1988. “Rhetorical Structure Theory: Toward a Functional Theory of Text Organization.” Text 8 (3): 243–281.

McEnery, Tony, and Andrew Hardie. 2012. Corpus Linguistics: Method, Theory and Practice. Cambridge: Cambridge University Press.

Menocal, María Rosa. 2002. The Ornament of the World: How Muslims, Jews, and Christians Created a Culture of Tolerance in Medieval Spain. Boston: Little, Brown.

Mitchell, Margaret. 2021. “Bias and Fairness in Natural Language Processing.” AI Magazine 42 (3): 36–52.

Mohammad, Saif M., and Peter D. Turney. 2013. “Crowdsourcing a Word–Emotion Association Lexicon.” Computational Intelligence 29 (3): 436–465.

Moretti, Franco. 2013. Distant Reading. London: Verso.

Papineni, Kishore, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. “BLEU: A Method for Automatic Evaluation of Machine Translation.” In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics (ACL), 311–318.

Pennebaker, James W., Martha E. Francis, and Roger J. Booth. 2001. Linguistic Inquiry and Word Count: LIWC 2001. Mahwah, NJ: Lawrence Erlbaum Associates.

Popović, M., Toral, A., & Sánchez-Cartagena, V. 2016. Impact of translationese on machine translation performance. Proceedings of the 19th Annual Conference of the European Association for Machine Translation (EAMT 2016).

Ramsay, Stephen. 2011. Reading Machines: Toward an Algorithmic Criticism. Urbana: University of Illinois Press.

Ryan, Marie-Laure. 2004. Narrative as Virtual Reality. Baltimore: Johns Hopkins University Press.

Santoyo, Julio-César. 2009. A través del espejo: estudios de traducción en la Edad Media. León: Universidad de León.

Schneider, B. 2022. Multilingualism and AI: The regimentation of language in the age of digital capitalism. Signs and Society, 10(3), 362–387. https://doi.org/10.1086/721757

Shutova, Ekaterina. 2010. “Models of Metaphor in NLP.” In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics (ACL).

Spivak, G. C. 2000. "The politics of translation". In L. Venuti (Ed.), The Translation Studies Reader (pp. 397–416). Routledge.

Stamatatos, Efstathios. 2009. “A Survey of Modern Authorship Attribution Methods.” Journal of the American Society for Information Science and Technology 60 (3): 538–556.

Toury, Gideon. 1995. Descriptive Translation Studies and Beyond. Amsterdam: John Benjamins.

Torres, C., & Ramos, I. 2018. Anglicisms in Portuguese translations: Evidence from EU institutional texts. Meta, 63(2), 456–476.

van Dijk, Teun A. 2008. Discourse and Context: A Sociocognitive Approach. Cambridge: Cambridge University Press.

Vanmassenhove, E., & Shterionov, D. 2019. Lost in translation: Loss of lexical richness in neural machine translation. Proceedings of Machine Translation Summit XVII, Dublin.

Venuti, Lawrence. 1995. The Translator’s Invisibility: A History of Translation. London: Routledge.

Vieira, L. N. 2023. Translation infrastructures in the age of AI. Translation Studies, 16(1), 1–18.

12. Appendix: Citation Formats

If you use or reference this framework, please cite:

@software{helton_nfaf_2025,
  author       = {Helton, Johennie},
  title        = {NFAF: Narrative Flattening Analysis Framework (v0.4.1)},
  year         = 2025,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.16885076},
  url          = {https://doi.org/10.5281/zenodo.16885076}
}