💬 NLP

Hybrid topic modelling for computational close reading: Mapping narrative themes in Pushkin's Evgenij Onegin

This study proposes a hybrid computational framework combining Latent Dirichlet Allocation and sparse Partial Least Squares Discriminant Analysis to perform a reproducible, close-reading analysis of narrative themes in Pushkin's *Evgenij Onegin*, demonstrating how lightweight probabilistic models can effectively map thematic structures in small-corpus literary texts despite relying on a single translated version.

Original authors: Angelo Maria Sabatini

Published 2026-03-23

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Angelo Maria Sabatini

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to understand a massive, intricate tapestry woven from thousands of threads. This tapestry is a famous Russian poem, Eugene Onegin, written by Pushkin. It's a story about love, boredom, duels, and fate, told in verse.

Traditionally, a human reader would sit down with a cup of tea and read the poem slowly, page by page, to understand its themes. But what if you wanted to see the "big picture" of the story's emotional landscape all at once? What if you wanted to map out exactly where the sadness, the romance, and the social drama happen, without reading every single word yourself?

This paper is about building a digital microscope to do exactly that. The author, Angelo Maria Sabatini, created a new method called "Hybrid Topic Modelling" to perform a "computational close reading" of the poem.

Here is how it works, explained through simple analogies:

1. The Problem: Too Many Threads, Too Few Eyes

The poem is long and complex. If you just throw all the words into a blender (a standard computer method called "bag-of-words"), you lose the story's shape. The computer sees "love," "snow," and "duel" as just random words next to each other, missing the fact that "snow" sets a sad mood while "duel" sets a violent one.

The author wanted to find the hidden themes (like "Romantic Longing" or "Social Pretense") that run through the story, but he needed a way to make sure the computer wasn't just guessing.

2. The Solution: A Two-Step Detective Team

The author didn't use just one tool; he used a team of two detectives with different skill sets.

Detective A (The Unsupervised Observer - LDA):
Imagine a detective who walks into a room full of people chatting and just listens without asking questions. They look for patterns: "Hey, every time the word 'snow' appears, the words 'winter' and 'loneliness' show up too."
This is Latent Dirichlet Allocation (LDA). It groups words that frequently appear together to guess what the "topic" is. It's great at finding patterns, but sometimes it gets a little fuzzy.
Detective B (The Supervised Interrogator - sPLS-DA):
Now, imagine a second detective who is very strict. They say, "Okay, Detective A, you think this group of words is about 'Sadness.' Let me test that. I'm going to look only at the words that prove it's 'Sadness' and ignore everything else."
This is sPLS-DA. It acts as a strict filter. It takes the groups Detective A found and asks, "Are these words really unique to this group, or are they just common words that show up everywhere?"

The Magic: By combining them, the author gets the best of both worlds. Detective A finds the broad themes, and Detective B sharpens the focus, ensuring the themes are real and distinct.

3. The Experiment: Mapping the Story

The author took the Italian translation of Eugene Onegin and chopped it up into 35 chunks (like chapters in a comic book). He fed these chunks into his "Detective Team."

What did they find?
The computer identified 5 main themes that act like the "emotional weather" of the story:

Nature & Lyrical Imagery: The cold, snowy, lonely Russian winter. This is Tatiana's world (the quiet, book-loving girl).
Individual Dreams: The big questions about life, glory, and poetry. This is often the narrator or the poet Lensky.
Emotional Overflow: The intense, messy feelings of love, heartbreak, and passion. This is the core of the love letters.
Social Noise: The balls, the gossip, the fashion, and the fake smiles of high society. This is Eugene's world (the bored aristocrat).
Conflict & Honor: The tension, the duels, the blood, and the strict rules of society that lead to tragedy.

4. The "Narrative Hubs": Checking the Map

To make sure the computer wasn't hallucinating, the author looked at specific, famous scenes in the book (called "Narrative Hubs").

Scene 1: Tatiana's Love Letter. The computer correctly identified this as a mix of "Emotional Overflow" and "Nature." It saw words like heart, dream, silence, and snow.
Scene 2: Eugene's Reply. The computer saw the same words, but the "flavor" changed. It was still emotional, but now mixed with "Social Rules" and "Distance." Eugene is trying to be cool and detached, and the computer caught that shift.
Scene 3: The Final Rejection. Years later, when they meet again, the computer saw that Tatiana had shifted from "Nature" to "Social Rules." She is now a married lady who must follow society's rules, even though she still loves him.

5. Why This Matters

Usually, computers are used to read millions of books to find trends (Distant Reading). But here, the author used the computer to read one book very closely (Computational Close Reading).

The Analogy:
Think of reading a poem like walking through a garden.

Traditional Reading: You walk the path, smelling every flower, noticing the texture of the leaves.
Standard Computer Analysis: You take a drone photo of the whole garden from 10,000 feet. You see green and brown, but you miss the details.
This Paper's Method: It's like having a drone that can zoom in on specific patches of flowers, identify exactly which species are growing together, and then draw a map showing you that the "sad blue flowers" always grow near the "stormy rocks," while the "happy yellow flowers" grow near the "sunlight."

The Takeaway

The paper proves that even with a small amount of text (just one poem), we can use math to create a reproducible map of a story's soul. It doesn't replace the human reader; instead, it gives the reader a new pair of glasses to see the hidden patterns of emotion and structure that we might miss when we are too close to the text.

It shows that Eugene Onegin isn't just a story about a guy and a girl; it's a structured journey where the "weather" of the story shifts from the cold isolation of nature to the noisy chaos of society, and finally to the tragic collision of the two. The computer helped us see that map clearly.

1. Problem Statement

The paper addresses the challenge of applying topic modelling to single, high-density literary texts (specifically poetry) rather than large-scale corpora.

Limitations of Standard Approaches: Traditional Latent Dirichlet Allocation (LDA) is designed for "distant reading" of massive datasets. When applied to short texts or single works, it suffers from data sparsity, instability due to small sample sizes, and difficulty in capturing narrative dynamics.
The "Bag-of-Words" Issue: Poetry relies on rhythm, sound, and figurative language, which often violates the bag-of-words assumption inherent in standard topic models, leading to opaque or uninterpretable results.
The Goal: The study aims to develop a hybrid framework that combines unsupervised topic discovery with supervised validation to perform "computational close reading." The objective is to map thematic structures and their longitudinal evolution within a single, well-defined work (Evgenij Onegin) to support, rather than replace, traditional literary interpretation.

2. Methodology

The study employs a hybrid workflow integrating LDA (unsupervised) and sparse Partial Least Squares Discriminant Analysis (sPLS-DA) (supervised).

A. Corpus and Preprocessing

Source: An Italian translation of Puškin's Evgenij Onegin by Giuseppe Ghini (2021).
Segmentation: The text was segmented into 35 documents (blocks), where each block consists of 10–11 contiguous stanzas. This aggregation was necessary to provide sufficient lexical signal for LDA, avoiding the instability of analyzing individual stanzas.
Lemmatization & Filtering:
- Used UDPipe for tokenization and part-of-speech (PoS) tagging.
- Vocabulary Selection: Retained Nouns, Proper Nouns, and Adjectives. Verbs were explicitly excluded because their high frequency tended to foreground narrative progression rather than stable thematic fields.
- Stop-words: A minimal, manually curated list was used to preserve stylistic nuances.
- Normalization: Proper names were normalized (e.g., handling diminutives) to ensure consistent character tracking without dominating topics.

B. Unsupervised Modelling (LDA)

Model: Latent Dirichlet Allocation (LDA) using a collapsed Gibbs sampler (topicmodels R package).
Parameters:
- Topics ( $K$ ): Fixed at 5 based on stability and interpretability tests (comparing $K=3, 5, 8$ ).
- Priors: $\alpha = 2$ (document-topic) and $\beta = 0.15$ (topic-term). These were chosen specifically for small corpora to encourage mixed topic distributions and smooth rare terms, deviating from standard large-corpus defaults.
Stability Protocol: To address the stochastic nature of Gibbs sampling and "label switching," the authors ran 100 independent chains. A multi-seed consensus protocol was used:
- Runs were aligned to a reference run (maximizing log-likelihood) using the Hungarian algorithm.
- Alignment criteria included Jaccard overlap of top terms and Spearman correlation of document-topic mixtures.
- Only runs meeting strict stability thresholds were averaged to produce a consensus topic distribution.

C. Supervised Validation (sPLS-DA)

Purpose: To act as a "supervised probe" to validate the LDA-derived topics and identify discriminant lexical markers.
Approach:
- The dominant LDA topic assigned to each document served as the ground truth label.
- One-vs-Rest (OvR): Five binary classifiers were trained to distinguish each topic from the others.
- Parameters: $ncomp = 2$ latent components; $keepX = 30$ variables (terms) per component to enforce sparsity.
- Validation: Stratified 5-fold cross-validation (repeated 5 times) was used to assess accuracy against random baselines.
Lexicon Consolidation: A final multiclass sPLS-DA model was fitted to rank terms by selection frequency and loading magnitude. Terms were classified as core-exclusive (specific to one theme) or shared (bridging themes) based on an exclusivity index ( $E$ ) and loading magnitude ( $|L|$ ).

D. Narrative Hubs

To bridge the gap between statistical blocks and literary episodes, the authors defined 13 Narrative Hubs (contiguous stanza groups corresponding to key plot points, e.g., "Tat'jana's letter," "The duel").
Hub-level topic mixtures were calculated as weighted averages of the underlying block distributions, allowing for a granular analysis of specific narrative moments.

3. Key Results

A. Thematic Structure (5 Topics)

The model successfully identified five stable, interpretable themes:

Natural & Lyrical Imagery: Associated with Tat'jana, nature (winter, snow, forest), and folklore.
Individual Expectations & Aspirations: Associated with Romantic self-reflection, poetry, and existential crisis (Evgenij's early state).
Lyrical–Sentimental Effusion: The most affectively charged theme, focusing on love, heart, and dreams (dominant in epistolary scenes).
Social & Mundane Milieu: Focuses on balls, fashion, and social performance (Saint Petersburg scenes).
Feelings & Social Conventions: Centers on conflict, honor codes, and the duel (violence vs. propriety).

B. Model Performance & Stability

Stability: The consensus protocol yielded high stability (Mean Jaccard@30 = 0.41; Mean Spearman correlation = 0.91).
Effective Number of Topics ( $N_{eff}$ ): Calculated as 4.52, indicating a balanced distribution where all 5 topics are well-represented.
Supervised Validation: The sPLS-DA classifiers achieved 85% overall accuracy and 70% balanced accuracy, significantly outperforming random baselines. This confirmed that the LDA-derived topics represent robust lexical structures rather than noise.
Convergence: A heatmap analysis showed strong diagonal agreement between LDA top terms and sPLS-DA discriminant terms, validating the thematic coherence.

C. Narrative Dynamics

Longitudinal Mapping: The topic mixtures tracked the narrative arc, showing how themes shift from "Natural" (Chapter 1) to "Social" (Chapter 8).
Epistolary Analysis: The study analyzed the four letter-writing scenes (H5–H6 and H12–H13).
- Tat'jana's Letter: Dominated by Lyrical–Sentimental Effusion and Natural Imagery.
- Evgenij's Reply: Retained Lyrical–Sentimental elements but shifted toward Individual Expectations, reflecting his emotional distance.
- Role Reversal: In the final chapters, Evgenij's letter showed intrusion of Feelings & Social Conventions (honor codes), while Tat'jana's final reply shifted toward Individual Expectations, signaling her transformation into a socially codified figure.

4. Key Contributions

Hybrid Framework for Close Reading: Proposes a novel workflow combining unsupervised topic modelling with supervised discrimination to analyze single literary works, moving beyond "distant reading" to "computational close reading."
Small-Corpus Stability Protocol: Introduces a rigorous multi-seed consensus and alignment method to overcome the instability of LDA in small datasets ( $N=35$ documents).
Methodological Decisions for Poetry: Demonstrates the strategic exclusion of verbs and the inclusion of proper nouns to prioritize thematic interpretability over full lexical coverage in verse.
Narrative Hubs: Develops a method to map statistical topic distributions onto specific literary episodes (hubs), allowing for the quantitative analysis of plot dynamics and character evolution.
Validation of Interpretability: Proves that lightweight probabilistic models can yield reproducible thematic maps that align with established literary criticism, even when abstracting away prosodic features.

5. Significance

Bridging Quantitative and Qualitative: The study demonstrates that computational methods do not need to replace traditional literary interpretation but can serve as a transparent, reproducible tool to formalize and visualize thematic structures that readers intuitively perceive.
Applicability: The framework is presented as a template applicable to other high-density literary texts and comparative studies, particularly where stylistic features (meter, phonology) are secondary to lexical thematic structure.
Insight into Puškin: The analysis provides a data-driven confirmation of the novel's structural duality (nature vs. society, emotion vs. convention) and the tragic trajectory of its protagonists, highlighting how thematic mixtures align with the poem's emotional and structural arc.