[Paper Review] From Once Upon a Time to Happily Ever After: Tracking Emotions in Novels and Fairy Tales

Saif M. Mohammad|arXiv (Cornell University)|Sep 23, 2013

Sentiment Analysis and Opinion MiningComputer Science24 references183 citations

TL;DR

This paper introduces an emotion-based text analysis framework using the NRC Emotion Lexicon to quantify and visualize emotion word densities in literary texts. It demonstrates that fairy tales exhibit significantly wider distributions of emotion word densities—especially for joy, surprise, anticipation, and disgust—compared to novels, enabling new forms of affective search and comparative literary analysis across large corpora like Google Books and Project Gutenberg.

ABSTRACT

Today we have access to unprecedented amounts of literary texts. However, search still relies heavily on key words. In this paper, we show how sentiment analysis can be used in tandem with effective visualizations to quantify and track emotions in both individual books and across very large collections. We introduce the concept of emotion word density, and using the Brothers Grimm fairy tales as example, we show how collections of text can be organized for better search. Using the Google Books Corpus we show how to determine an entity's emotion associations from co-occurring words. Finally, we compare emotion words in fairy tales and novels, to show that fairy tales have a much wider range of emotion word densities than novels.

Motivation & Objective

• To develop a sentiment and emotion analysis system for large-scale literary text collections.
• To enable emotion-based search in digitized literature, such as finding texts with high suspense or joy.
• To compare emotional dynamics between fairy tales and novels using quantitative emotion word density metrics.
• To visualize and analyze emotion distributions across genres and authors for social and stylistic insights.
• To create an affect-based interface for Project Gutenberg, enhancing access to literary texts through emotional content.

Proposed method

• The study employs the NRC Emotion Lexicon, a crowdsourced word–emotion association lexicon, to assign emotion labels to words in texts.
• Emotion word density is calculated as the number of emotion-labeled words per 10,000 words in a text, enabling cross-text comparison.
• Visualizations such as histograms and density plots are used to compare emotion distributions across corpora, including novels and fairy tales.
• The Google Books Corpus is leveraged to analyze entity-specific emotion associations through co-occurrence patterns with emotion words.
• Statistical tests (p < 0.001) are applied to assess significance in differences between emotion word densities in novels and fairy tales.
• The analysis is applied to two corpora: the CEN (292 novels, 1881–1922) and the FTC (453 fairy tales, 19th-century authors like Grimm, Andersen, Potter).

Experimental results

Research questions

• RQ1 How do emotion word densities differ between fairy tales and novels in terms of distribution and magnitude?
• RQ2 Can emotion-based visualizations improve search and exploration of large literary corpora?
• RQ3 Do entities such as women, race, or homosexuals show distinct emotion word associations in historical texts?
• RQ4 How do emotion word densities vary across different emotions (e.g., joy, fear, surprise) in fairy tales versus novels?
• RQ5 Is there a significant difference in the standard deviation of emotion word densities between fairy tales and novels?

Key findings

• Fairy tales have significantly higher densities of joy, surprise, anticipation, and disgust compared to novels (p < 0.001).
• Fairy tales exhibit significantly lower trust word density than novels (p < 0.001).
• Fairy tales have a much larger standard deviation in emotion word densities across all eight basic emotions compared to novels (p < 0.001), indicating greater emotional variability.
• For every 10,000 words, novels contain an average of 1,670 negative and 2,602 positive words, while fairy tales contain 1,543 negative and 2,808 positive words (p < 0.001).
• The distribution of emotion word densities in fairy tales is not bimodal but approaches a normal distribution with more extreme values than novels, indicating broader emotional range.
• Visualizations show that fairy tales have more texts with both very high and very low emotion word densities than novels, confirming wider emotional variation.