A RoBERTa-large model finetuned for 3-class sentiment classification (negative / neutral / positive) on literary and fictional text. It is designed for sentence-level sentiment scoring of novels, short stories, and other narrative prose.

Model description

This model is a finetuned version of j-hartmann/sentiment-roberta-large-english-3-classes (RoBERTa-large, 355M parameters). It was trained on a combined corpus of human-annotated fiction sentences from multiple sources, using class-weighted cross-entropy loss to handle label imbalance.

Training data

Only human-annotated texts.

Source	n (train)	Label type
Project Gutenberg and Wattpad excerpts	6,646	Nine emotions labels → binned to 3 classes
EmoBank Fiction (American National Corpus)	2,164	Continuous valence → binned to 3 classes
Fiction4 Hymns (translated from Danish)	1,620	Continuous valence → binned to 3 classes
Hemingway — The Old Man and the Sea	1,554	Continuous 1–10 valence → binned to 3 classes
Fiction4 Poetry (Plath)	1,263	Continuous valence → binned to 3 classes
Fiction4 Fairy Tales (Andersen, translated)	617	Continuous valence → binned to 3 classes
Total	13,864

Continuous valence scores were binned using the thresholds: ≤4 → negative, (4, 6] → neutral, >6 → positive on a 0–10 scale.

Intended use

This model is intended for research on literary sentiment, narrative emotion arcs, and computational literary studies. It can be used for:

Sentence-level sentiment classification of fiction and literary prose
Generating continuous sentiment arcs by converting class probabilities to a valence score: valence = p(positive) - p(negative)
Comparing sentiment patterns across genres, authors, or narrative structures

Evaluation

All evaluation sets were held out from training. Continuous valence for correlation is computed as p(positive) − p(negative) from the model's softmax probabilities, yielding a score in approximately [−1, +1] rather than a discrete class label. Spearman ρ is computed against continuous human valence annotations where available, or against ordinal 3-class labels. Accuracy is computed on the 3-class prediction (argmax over negative/neutral/positive) against human valence binned with the same thresholds used for training (≤4 → negative, (4, 6] → neutral, >6 → positive). Note that literary texts are heavily neutral-skewed, where always predicting "neutral" would do better. For this reason, the continuous valence correlation (Spearman ρ) is the more meaningful metric here.

Eval set	n	Spearman ρ	Pearson r	Accuracy	Baseline (Syuzhet)
Hemingway test	187	0.714	0.729	0.845	0.307
Book passages test	839	0.754	0.759	0.782	0.578
EmoBank Fiction	271	0.754	0.785	0.804	0.517
Fiction4 Poetry (Plath)	158	0.723	0.768	0.791	0.473
Fiction4 Fairy Tales (Andersen)	78	0.674	0.743	0.705	0.611
Fiction4 Hymns	203	0.821	0.801	0.739	0.630

The Hemingway inter-annotator agreement (Spearman ρ between two human annotators) is 0.543, which the model substantially exceeds on the held-out test set.

The Syuzhet baseline is a dictionary-based method using the Syuzhet lexicon (Jockers, 2015).

Comparison with base model (v2)

The base model (v2, not released) was finetuned only on Gutenberg and Wattpad passages + Hemingway (8,200 training sentences). This model (v3) adds EmoBank Fiction and Fiction4 subsets (13,864 training sentences).

Eval set	v3 Spearman ρ	v2 Spearman ρ	Δ
Hemingway test	0.714	0.655	+0.059
EmoBank Fiction	0.754	0.701	+0.053
Fiction4 Poetry	0.723	0.652	+0.070
Fiction4 Hymns	0.821	0.785	+0.036
Fiction4 Fairy Tales	0.674	0.681	−0.007
Books test	0.754	0.780	−0.025

v3 improves on literary/fiction benchmarks with continuous human annotations. The slight drop on Books test (excerpts with ordinal labels) reflects a trade-off from the more diverse training mix.

Sequential arc variant: sentiment-fiction-seq

A variant of this model, fpianz/sentiment-fiction-seq, was trained on a modified split where complete sequential texts were held out from training to evaluate detrended sentiment arcs. The training data is slightly reduced (12,929 sentences) because all three Andersen fairy tales and a contiguous block of 400 Hemingway sentences were removed from training and reserved for sequential evaluation.

Detrending follows the nonlinear adaptive filtering method of Hu et al. (2021): the sentiment arc is integrated into a random walk, an adaptive polynomial filter extracts the global trend, and the residuals capture local narrative dynamics. Spearman ρ is computed between the detrended model arc and the detrended human annotation arc.

Eval set	n	Raw Spearman ρ (Tr)	Raw Spearman ρ (Sy)	Detrended Spearman ρ (Tr)	Detrended Spearman ρ (Sy)
Hemingway — The Old Man and the Sea	400	0.712	0.465	0.781	0.335
Andersen — The Ugly Duckling	211	0.600	0.469	0.741	0.584
Andersen — The Little Mermaid	293	0.654	0.523	0.754	0.624
Andersen — The Shadow	267	0.734	0.456	0.796	0.657

Tr = Transformer, Sy = Syuzhet lexicon baseline. Detrended values use adaptive filter at window L/8. The Hemingway inter-annotator agreement is Spearman ρ = 0.613.

Detrending consistently improves the transformer's correlation with human annotations (+0.06 to +0.14), indicating that the model captures arc-level narrative dynamics beyond sentence-level sentiment. These results are limited to Andersen fairy tales (translated from Danish) and one Hemingway novella. Users working with other types of fictional text should validate on their own data to determine which model variant (sentiment-fiction or sentiment-fiction-seq) best fits their use case.

Usage

python

from transformers import pipeline

classifier = pipeline("text-classification", model="fpianz/sentiment-fiction")
result = classifier("The old man was thin and gaunt with deep wrinkles in the back of his neck.")
print(result)
# [{'label': 'negative', 'score': 0.82}]

For continuous sentiment arcs:

python

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained("fpianz/sentiment-fiction")
model = AutoModelForSequenceClassification.from_pretrained("fpianz/sentiment-fiction")

def valence(text):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
    with torch.no_grad():
        logits = model(**inputs).logits
    probs = torch.softmax(logits, dim=-1)[0]
    return (probs[2] - probs[0]).item()  # p(positive) - p(negative)

score = valence("He was an old man who fished alone in a skiff in the Gulf Stream.")
print(f"Valence: {score:.3f}")  # range approx [-1, +1]

Training details

Base model: j-hartmann/sentiment-roberta-large-english-3-classes
Architecture: RoBERTa-large (355M parameters)
Loss: Class-weighted cross-entropy (weights: negative=1.01, neutral=0.72, positive=1.60)
Epochs: 5 (with early stopping, patience=3)
Learning rate: 2e-5
Batch size: 16
Max sequence length: 512
Optimizer: AdamW (weight decay=0.01, warmup ratio=0.1)
Precision: FP16
Hardware: NVIDIA A100 (University of Groningen Habrok HPC)

Limitations

Fiction4 Fairy Tales and Hymns are Google-translated from Danish (Feldkamp et al., 2024); translation artifacts may affect those evaluation scores.
The 3-class label scheme (negative/neutral/positive) collapses the valence spectrum. The continuous valence conversion (p(pos) - p(neg)) provides finer granularity but is an approximation.
Hemingway sentences constitute ~11% of training data. Evaluation on Hemingway test (held out) is uncontaminated, but the model may be biased toward Hemingway's style.

References

Sentiment Below the Surface: Omissive and Evocative Strategies in Literature and Beyond (Feldkamp et al., CHR 2024)
DENS: A Dataset for Multi-class Emotion Analysis (Liu et al., EMNLP-IJCNLP 2019)
Comparing Tools for Sentiment Analysis of Danish Literature from Hymns to Fairy Tales: Low-Resource Language and Domain Challenges (Feldkamp et al., WASSA 2024)
Dynamic evolution of sentiments in Never Let Me Go: Insights from multifractal theory and its implications for literary analysis (Hu et al., DSH 2021)

Citation

Paper under review — citation will be added upon publication.