PaperlayerView exact revision on arXiv

Source-linked AI summary

LITERARYBIGFIVE: Author-Personalized Text Generation in a Unified Interpretable Space

Jinghui Zhang, Lang Gao, Ao Li, Mingzhe Li, Ruihong Zeng, Zirui Song, Kentaro Inui, Xiuying Chen

arXiv:2608.23124v1cs.CLcs.AI

TL;DR

Personalized literary generation is constrained by author-isolated representations that require costly author-specific data or retraining and offer limited interpretability. LiteraryBigFive instead derives five interpretable stylistic axes from author-versus-neutral activation contrasts, then localizes authors and steers generation toward target coordinates. Experiments report improved authorial expressiveness with preserved semantic fidelity, while coordinates align with established literary consensus.

  • Problem

    Existing author-personalization approaches treat authors as isolated categories, making adaptation costly and limiting unified interpretability across writing patterns.

  • Method

    LiteraryBigFive constructs a unified five-dimensional space from activation contrasts between author-written and neutral passages, then uses localization and steering for personalized generation.

  • Results

    LiteraryBigFive improves authorial expressiveness while preserving semantic fidelity, and its derived coordinates align with established literary consensus.

  • Takeaways & Limitations

    The framework provides a unified, interpretable representation of authorial characteristics with per-dimension explanations for personalized generation.

  • Takeaways & Limitations

    The axes are primarily derived from English literary classics, and global residual-stream steering lacks granularity for manipulating specific long-range dependencies.

Abstract

from arXiv · show

Personalized text generation for authors and literary writing is essential for applications such as adaptive writing assistants, creative support tools, and computational literary analysis. However, existing approaches to author modeling and personalization often represent writing behavior as independent labels, requiring large-scale corpus collection or fine-tuning for each author or stylistic category. Such formulations are costly, difficult to interpret, and poorly suited for generalizing across authors. Inspired by the Big Five model's dimensional view of personality, we propose LiteraryBigFive, a framework that reframes authorial writing characteristics as coordinates within a unified and interpretable space. In this space, we derive each interpretable axis (e.g., Classicism, Emotionality) from activation-space contrasts between author-written and neutral passages, yielding distinct stylistic dimensions that allow texts or authors to be positioned within a five-dimensional system. Beyond localizing different authors, we further introduce an interpretable steering mechanism, which adaptively guides text generation toward target coordinates to perform author-personalized writing. Experimental results show that LiteraryBigFive improves authorial expressiveness while preserving semantic fidelity. The derived author per-axis scores strongly correlate with real-world literary consensus, offering transparent and interpretable explanations of author-specific generation behavior: https://github.com/Znull-1220/LiteraryBigFive.

1 Introduction

Personalized literary generation is valuable for matching writing preferences and emulating authorial voices, but author-isolated approaches are costly and provide limited interpretability. LITERARYBIGFIVE addresses this by representing authorial characteristics in a unified five-dimensional space and using localization and steering for personalized generation.

  • Personalized text generation supports stylistic preference matching and voice emulation in creative writing and personalized assistants.
  • Existing methods model authors as isolated categories, requiring author-specific prompting, training, or steering.Adapting to a new author typically requires hundreds or thousands of texts or model retraining.
  • Linguistic and literary studies describe writing variation through stable, interpretable dimensions such as narrativity, emotion, and elaboration.
  • LITERARYBIGFIVE represents authorial characteristics as coordinates in a unified five-dimensional space rather than unrelated labels.Its axes are Classicism, Ornateness, Narrativity, Emotionality, and Analyticity; axis directions come from contrasts between author-written and neutral passages, followed by axis decomposition to improve independence.
  • The localize-and-steer mechanism positions authors or books using shared coordinates and steers model activations toward target authorial coordinates for personalized generation.
  • Experiments on books with distinct writing identities show better target-author matching while preserving meaning, and learned coordinates align with established literary consensus.

2 Related Work

Prior work studies personalized generation through supervised rewriting and unsupervised disentanglement, while dimensional models have been used for personality assessment and persona simulation. Controllable personalized text generation remains underexplored from this dimensional perspective.

  • Personalized generation traditionally uses supervised rewriting with parallel corpora or unsupervised disentanglement to separate content from linguistic expression.
  • Psychological Big Five frameworks describe individual variation through compact interpretable axes, which NLP has applied to personality assessment and persona simulation.
  • Categorical style or author labels still dominate controllable personalized text generation, leaving the dimensional perspective underexplored.

3 Problem Formulation

LITERARYBIGFIVE formulates author-personalized generation as coupled localization and steering tasks in a unified five-dimensional space. Localization estimates an author’s position from reference passages, while steering rewrites neutral text toward a target position while preserving semantics.

  • Author-personalized generation is formulated as two coupled sub-tasks: localization and steering in a unified five-dimensional space.
  • A locator maps reference passages from a target book to coordinates in the LITERARYBIGFIVE space.
  • Averaging passage coordinates estimates the target author’s or book’s characteristic position within the space.
  • Steering takes a neutral input passage and target position as inputs for generating a rewritten passage.
  • The rewritten passage is intended to preserve the neutral input’s semantics while aligning its linguistic expression with the target position.

4 Method

LITERARYBIGFIVE constructs five interpretable literary-style axes from paired author-written and neutral passages, then localizes books and steers generation in that shared space. Axis refinement removes a shared expressiveness trend, while projections produce calibrated authorial targets for controllable rewriting.

  • Space construction: The framework defines five axes—Classicism, Ornateness, Narrativity, Emotionality, and Analyticity—to represent major variations in English literary writing.The dimensions cover traditional literary patterns, lexical and syntactic complexity, storytelling, affective intensity, and reasoning-oriented exposition.
  • Space construction: Representative literary passages are paired with semantics-preserving neutral rewrites to isolate authorial expression from content.The paired data are constructed from selected books, with authorial cues suppressed in the neutral versions.
  • Space construction: Axis directions are extracted from hidden-state contrasts between author-written and neutral passages, averaged across pairs, and renormalized.The contrast is designed to capture stylistic variation rather than semantic content or positional bias.
  • Axis refinement: A shared expressiveness direction is identified with PCA and removed from the raw axes before restoring layer-wise intervention scales.The decomposition reduces cross-axis correlations and supports more stable combinations of multiple dimensions.
  • Personalized steering: Generation computes the gap between current token scores and target authorial coordinates, then updates hidden states along interpretable axes.The approach enables explicit per-dimension modulation while preserving the input passage’s semantic content.

5 Experiments

The experiments evaluate LITERARYBIGFIVE on diverse literary books using automatic, LLM-based, and human measures of reconstruction, semantic preservation, and authorial adherence. Across four books, the framework consistently outperforms the compared baselines.

  • Evaluation Setup: The evaluation uses 590 passage-level samples from four books by Burke, Orwell, Stevenson, and Austen.The dataset totals 5,716 sentences and spans distinct writing identities.
  • Evaluation Setup: The study measures reconstruction quality with ROUGE-1/L and semantic preservation with embedding similarity.Embedding similarity uses cosine similarity between BGE sentence representations.
  • Evaluation Setup: GPT-4 and human annotators evaluate Authorial Adherence and Semantic Fidelity using the same two-dimensional criteria.GPT-4 rates passages on a 0–10 scale, while two annotators provide complementary human evaluations.
  • Compared Methods: LITERARYBIGFIVE is compared with few-shot prompting, supervised fine-tuning, and activation-steering baselines on a Llama2-7B-Chat base model.Additional results use Qwen2.5-3B-Instruct, while the main experiments use NVIDIA RTX 5880 Ada GPUs.
  • Main Results: Across four stylistically diverse books, LITERARYBIGFIVE consistently outperforms all baselines on ROUGE, SIM, GPT-4, and human evaluations.Table 1 scales all results to 0–100 and reports two-tailed paired t-tests with p<0.01.
  • Main Results: The framework maintains high scores across domains while improving authorial adherence and preserving semantic fidelity.The reported evaluation emphasizes robustness across authors and alignment with target-author characteristics without altering original meaning.

6 Analysis and Discussion

The analysis tests whether the framework’s components support personalization and whether its coordinates correspond to interpretable stylistic judgments. Results support adaptive axis-based steering, semantic preservation, and strong agreement with LLM-based authorial assessments.

  • 6.1 Ablation Study: Removing axis decomposition causes a noticeable drop across all metrics, showing that refinement isolates dimension-specific authorial signals.The ablation compares the full method with raw book-level directions.
  • 6.1 Ablation Study: Disabling dynamic style-gap adaptation causes an even larger performance degradation than removing refinement.The fixed steering-strength variant performs worse than the variant without axis decomposition.
  • 6.1 Ablation Study: Adaptive token-level steering is necessary because authorial cues are unevenly distributed across passages and require context-sensitive adjustment.The analysis links dynamic intervention to avoiding insufficient or excessive steering.
  • 6.2 Authorial Coordinates Analysis: The study compares model-derived coordinates with independent stylistic judgments from frontier LLMs.GPT-5, Claude 3.5, and Gemini 3 rate books on the five LITERARYBIGFIVE dimensions.
  • 6.2 Authorial Coordinates Analysis: The model coordinates achieve an average Pearson correlation of r = 0.96 with the ensemble LLM scores across all axes.The radar charts also reflect high Classicism for Edmund Burke and high Analyticity for George Orwell in both assessments.
  • 6.3 Case Study: Dimension Steering: Steering individual axes with strengths α ∈{−0.8, −0.4, 0, +0.4, +0.8} modulates corresponding stylistic dimensions without changing underlying meaning.The case study isolates one target dimension at a time while keeping the others at zero; an example is the shift toward Latinate diction under Classicism.

7 Conclusion

LITERARYBIGFIVE unifies authorial characteristics in an interpretable five-dimensional space and combines coordinate analysis with low-cost personalized generation. Experiments report stronger authorial expressiveness and semantic fidelity than baselines, with coordinates closely matching literary consensus.

  • Conclusion: LITERARYBIGFIVE reframes isolated authorial writing characteristics as coordinates in a unified and interpretable five-dimensional space.The framework replaces isolated author labels with shared stylistic dimensions.
  • Conclusion: Its localize-and-steer mechanism supports precise authorial analysis and low-cost personalized generation for new authors.The conclusion describes the approach as adapting to new authors without requiring the same kind of author-specific retraining emphasized earlier.
  • Conclusion: Experimental results demonstrate improved authorial expressiveness and semantic fidelity, while derived coordinates closely match established literary consensus.The conclusion also identifies multilingual settings and interactive writing support as future directions.

Limitations

The framework is constrained by its English-literary basis, global residual-stream steering, and reliance on internal activation access. These boundaries limit linguistic coverage, fine-grained control, and compatibility with closed-source APIs.

  • Scope: The axes are primarily derived from English literary classics, limiting the study’s coverage of broader linguistic and literary settings.The authors identify extension to richer linguistic and literary settings as future work.
  • Steering Granularity: Global steering on residual-stream layers captures holistic writing attributes but lacks granularity for specific long-range dependencies.The authors suggest targeting individual attention heads or components as a possible way to address this limitation.
  • Model Access: Reliance on internal activation vectors requires white-box access, limiting applicability to open-weight models and excluding closed-source APIs without embedding access.This boundary follows from the framework’s extraction and manipulation of internal representations.

Ethical Considerations

The framework is developed using publicly available literary sources and an open, reproducible generation setting without private or personally identifiable data. Its procedure constructs literary axes and adaptively steers generation, while cross-model results indicate effectiveness beyond one backbone.

  • The datasets derive primarily from public-domain literary works and exclude private, sensitive, or personally identifiable information.
  • The framework procedure includes offline literary-space construction followed by online target localization and interpretable steering during generation.
  • LITERARYBIGFIVE constructs interpretable literary axes, localizes target author coordinates, and adaptively steers generation using the style gap between hidden states and targets.
  • On Qwen2.5-3B-Instruct, LITERARYBIGFIVE achieves the best performance across all four books and all evaluation metrics.The results suggest effectiveness is not tied to a specific backbone and extends across model series and scales.

C.1 Dataset Construction

The dataset is built from open literary classics, cleaned and paired with neutralized passages to isolate authorial expression from content. The resulting axes and analyses support separated, interpretable book-level coordinates and efficient steering.

  • The corpus uses open-access English literary classics, selecting 10 works by distinct well-known authors for axis construction and four held-out books for evaluation.The held-out books are Burke’s Reflections on the Revolution in France, Orwell’s 1984, Stevenson’s Kidnapped, and Austen’s Pride and Prejudice.
  • Author-written passages are paired with LLM-neutralized versions that preserve meaning while suppressing authorial cues, isolating voice from content differences.
  • A manual inspection of 100 passage pairs found them clean, meaning-preserving, and sufficiently suppressed in distinctive authorial expression for axis construction and evaluation.
  • The five axes cover Classicism, Ornateness, Narrativity, Emotionality, and Analyticity, with literary anchors grounded in established literary and linguistic analysis.Examples include action- and time-marker-based Narrativity and logic-driven Analyticity.
  • After decomposition, mean absolute offdiagonal cosine similarity decreases from 0.87 to 0.27, indicating more disentangled axis directions.
  • The 95th percentile of absolute anchor projections calibrates each axis, making coordinates comparable while limiting outlier influence and avoiding saturation.
  • Book-level score distributions separate consistently with known literary patterns, while the steering mechanism has linear per-token complexity O(K ·d).LITERARYBIGFIVE records 19.88 ms/token, less than 1.0 ms above the fastest static baseline and effectively equivalent to LLM-Steer at 19.92 ms/token.

I BIGFIVE Dimension Vector Analysis

Layer-wise probing finds all five literary dimensions robustly represented in the model, but at different depths. Surface dimensions emerge early, whereas semantic and affective dimensions require deeper processing.

  • All five dimensions achieve AUC > 0.90, indicating robust dimension-level information in the model’s latent space.
  • Figure 7 reports separated per-book distributions across the five dimensions that match known literary characteristics.
  • Classicism and Narrativity reach near-perfect separability within Layers 0–5, consistent with lexical and shallow syntactic cues.
  • Analyticity, Emotionality, and Ornateness peak in Layers 15–25, with Analyticity showing greater lower-layer variance.
  • The results indicate that deeper abstraction supports representations of reasoning processes and affective nuances, while shallow layers encode surface patterns.

J More Authorial Coordinates Analysis

Additional author analyses place diverse writing styles in the shared five-dimensional space in ways aligned with literary criticism. Qualitative comparisons further show how the method balances stylistic adherence with semantic preservation.

  • Qualitative evaluations compare Semantic Fidelity and Authorial Adherence across four books, with results scaled to 0–100.
  • The model localizes six additional authors in the five-dimensional framework, with coordinates aligning closely with established literary criticism.
  • Hemingway and Faulkner both show high Emotionality but differ in Ornateness by Δ=4.6, reflecting sparse versus multi-layered prose.
  • Compared with baselines, the method preserves rhetorical structure, specific details, narrative flow, and authorial voice in representative cases.
  • Few-shot prompting produces irrelevant context or hallucinated scenes in several examples, while RepE can preserve meaning but flatten tone or rhetorical structure.
Loading 2608.23124v1…