Modeling Multilayered Complexity in Literary Texts

Pascale Feldkamp, Márton Kardos, Kristoffer L. Nielbo, Yuri BizzoniNoDaLiDa/Baltic-HLT 2025 · OpenReview

TL;DR

This paper investigates how stylistic and sentiment complexity interact in literary texts, finding a strong linear association between them. Surprisingly, these two dimensions of complexity tend to co-occur rather than trade-off, suggesting they jointly contribute to a text's perceived difficulty and processing time.

What Problem Does This Paper Address?

Literary texts are known for their intricate complexity, engaging readers on multiple levels. Traditional metrics for text difficulty, often developed for non-fiction, struggle to capture this richness. While stylistic and syntactic complexity have been extensively studied, the role of sentiment and emotional complexity has been relatively overlooked. This paper aims to bridge this gap by exploring the relationship between stylistic/syntactic complexity and sentiment complexity, and how these combined factors influence reader experience, specifically cognitive load as proxied by reading time and difficulty rank.

Methods

The authors employ a multi-faceted approach:

  • Correlation Analysis: They first examine the correlation between a wide range of stylistic/syntactic features and sentiment complexity measures (sentiment standard deviation, approximate entropy, Hurst exponent) across a large corpus of novels.
  • Linear Regression: To understand predictive relationships, they build linear regression models. This includes predicting sentiment complexity from stylistic/syntactic features and vice-versa, as well as predicting reader experience proxies (reading time, difficulty rank) from these feature sets.
  • Recursive Feature Elimination (RFE): To identify the most important features for prediction in smaller datasets, RFE is used.
  • Principal Component Analysis (PCA): For the smaller datasets (reading time and difficulty rank), PCA is applied to reduce collinearity and improve the robustness of the regression models.

Datasets

  • Chicago Corpus: A large dataset of over 9,000 English novels from the 19th and 20th centuries, used for the primary correlation and regression analyses. This corpus is notable for its temporal breadth and selection based on library holdings, aiming to capture popular and prestigious literature.
  • Natural Stories Corpus: A smaller dataset of 10 English stories (approx. 1,000 words each) with associated self-paced reading (SPR) data from 181 native English speakers. This dataset is used to measure reading times.
  • Difficulty Rank List: A curated list of 200 novels, each assigned a difficulty rank based on a binary prediction task using linguistic and psycholinguistic features. This list was used to assess perceived difficulty for a subset of novels from the Chicago Corpus.

Key Insights and Ideas

  • Co-occurrence, Not Trade-off: Contrary to the "cognitive compensation" hypothesis (where complexity in one area might lead to simplification in another), the study finds a strong positive correlation between stylistic/syntactic complexity and sentiment complexity. This suggests that texts that are linguistically dense and complex also tend to have more varied and intricate sentiment arcs.
  • Sentiment Complexity Matters for Reader Experience: Sentiment features, particularly sentiment standard deviation, are shown to be significant predictors of both reading time and difficulty rank, mirroring the predictive power of stylistic/syntactic features. This highlights that emotional and semantic complexity is not a secondary concern but a crucial factor in how readers perceive and process literary texts.
  • "Supernatural Stimulus" Hypothesis: The findings support the idea that literary texts might function as "supernatural stimuli," where multiple elements are intensified simultaneously to create a more engaging and powerful experience. This is contrasted with nonfiction, where clarity and ease are often prioritized.
  • Nominal Style's Role: The study points to the intriguing role of "nominal style" (e.g., high frequency of "of," dense noun phrases) in literary texts. This style, often associated with efficient semantic information transfer in technical writing, might be adapted in literature to convey complex sentiment more effectively.

Key Results

  • Strong Correlations: Stylistic/syntactic features like readability, sentence length, and lexical richness show strong positive correlations with sentiment standard deviation and approximate entropy.
  • Predictive Power: Linear regression models demonstrate that stylistic/syntactic features can predict sentiment complexity, and vice-versa. For instance, sentiment standard deviation was a significant predictor of difficulty rank.
  • Reading Time & Difficulty: Sentiment features, alongside stylistic/syntactic ones, were found to be related to reading time and difficulty rank, suggesting they contribute to cognitive load. For example, sentiment SD was identified as a top significant feature in predicting reading time in one experiment.

Limitations and Future Work

  • Small Sample Sizes: The experiments involving reading time and difficulty rank were conducted on relatively small datasets (10 and 26 novels, respectively), which limited the scope of linear regression and potentially obscured some correlations.
  • Need for Larger Annotated Corpora: The authors suggest that a larger corpus of annotated fiction is needed to robustly confirm some of the observed tendencies, particularly regarding the non-linear correlations between features and reading time.
  • Future Directions: The authors plan to further explore the relationship between these complexity levels in literary language, better formalize their interplay, and conduct more experiments to test the relationship between perceived complexity and these features. They also aim to rigorously test differences between literary and non-literary texts regarding multidimensional complexity.

Summary updated

Abstract

We explore the relationship between stylistic and sentimental complexity in literary texts, analyzing how they interact and affect overall complexity. Using a dataset of over 9,000 English novels (19th-20th century), we find that complexity at the stylistic/syntactic and sentiment levels tend to show a linear association. Finally, using dedicated datasets, we show that both stylistic/syntactic features – particularly those relating to information density – as well as sentiment features are related to text difficulty rank as well as average processing time.
BibTeX
@inproceedings{feldkamp2025modeling,
  author = {Pascale Feldkamp and Márton Kardos and Kristoffer L. Nielbo and Yuri Bizzoni},
  title = {{Modeling Multilayered Complexity in Literary Texts}},
  booktitle = {NoDaLiDa/Baltic-HLT},
  year = {2025},
  url = {https://openreview.net/forum?id=eLy90JMOqk},
}
Figure