Source-linked AI summary
From Once Upon a Time to Happily Ever After: Tracking Emotions in Novels and Fairy Tales
Saif Mohammad
TL;DR
Digitized literary collections are abundant, but keyword-based search provides limited access to their emotional content. The paper combines sentiment analysis, visualizations, emotion word density, and corpus co-occurrence analysis to quantify emotions in books and collections. It finds that fairy tales have a much wider distribution of emotion word densities than novels.
Problem
Automated access to unprecedented amounts of digitized literary text still relies heavily on keyword searches, limiting emotion-oriented analysis and search.
Method
The paper uses a word–emotion association lexicon, a simple emotion analyzer, visualizations, emotion word density, and Google Books Corpus co-occurrence analysis.
Results
Fairy tales have a much wider distribution of emotion word densities than novels.
Takeaways & Limitations
Emotion analysis and visualization can support emotion-based search, tracking within books, corpus-scale entity associations, and comparison of literary collections.
Abstract
from arXiv · showhide
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.
1 Introduction
The paper addresses the limits of keyword-only access to digitized literary texts by combining sentiment analysis with visualizations to quantify and track emotions. It develops an emotion analyzer and applies emotion word density and corpus analysis to support searching, comparison, and analysis across books.
- Motivation: Digitization provides access to tens of thousands of books and more than 5.2 million digitized books, while automated access still relies heavily on keyword searches.Project Gutenberg provides 34,000 books, and the Google Books Corpus provides n-gram sequences and frequencies from more than 5.2 million books.
- Approach: The paper combines sentiment analysis with visualizations to quantify and track emotions in individual books and very large collections.The authors frame this approach as supporting emotion-based search, social analysis, literary comparison, summarization, and persuasion analysis.
- Approach: The authors use a large word–emotion association lexicon to create a simple emotion analyzer and visualize emotion-word use across texts and collections.The analyzer recognizes positive, negative, and neutral polarity and associates words with joy, sadness, anger, fear, trust, disgust, surprise, and anticipation.
- Applications: Emotion word density organizes the Brothers Grimm fairy tales for emotion-based search, while Google Books Corpus co-occurrences reveal emotion associations toward entities.The paper also compares emotion-word densities between novels and fairy tales.
- Findings: Fairy tales have a much wider distribution of emotion word densities than novels.This is the paper’s final comparative finding across the two literary collections.
2 Related work
Prior work established sentiment and emotion detection but often used smaller lexicons or human annotation that limited the scale of literary-text analysis. This paper instead analyzes individual texts and very large collections using out-of-context word–emotion associations.
- Sentiment analysis: Sentiment-analysis research has addressed positive and negative polarity, as well as emotions including anger, joy, sadness, fear, surprise, and disgust.The paper places its work within a developing body of sentiment-analysis research.
- Emotion lexicons: Existing resources included the small WordNet Affect Lexicon and the broader General Inquirer, which labels 11,788 words across 182 categories.The WordNet Affect Lexicon contains a few hundred words annotated for affect categories.
- Literary-text analysis: Human annotation has enabled empirical literary analysis but restricted the number of texts analyzed.One cited study annotated 22 Brothers Grimm fairy tales and found that tales often began neutrally and ended happily.
- This paper: The paper analyzes individual texts through very large collections using out-of-context word–emotion associations.Its approach relies on information from many words rather than human annotation of every text.
3 Emotion Analysis
The emotion-analysis framework builds on a crowdsourced word–emotion lexicon with word-sense filtering and majority-vote annotation. It then matches text words to lexicon entries to estimate emotion distributions over larger texts and corpora.
- Lexicon construction: The NRC Emotion Lexicon was crowdsourced and annotates words for polarity and eight emotions.The eight emotions are anger, fear, joy, sadness, disgust, surprise, trust, and anticipation.
- Lexicon construction: Only Roget’s Thesaurus words occurring more than 120,000 times in the Google n-gram corpus were annotated for the experimental lexicon version.The source terms came from the 1911 Roget Thesaurus.
- Word senses: Word-choice questions guide annotators toward the intended sense before they judge sentiment and emotion associations.Annotations for the remaining questions are discarded when the annotator answers the sense question incorrectly.
- Annotation: About 10% of annotations were discarded after incorrect word-choice responses, and the remaining annotations were combined by majority vote across five annotators.The lexicon contains about 24,200 word–sense pairs.
- Text analysis: For a target text, the analyzer matches words in the emotion lexicon and calculates ratios of emotion-associated words to total emotion words.The approach is not necessarily reliable for identifying emotion in a particular sentence, but is reliable for comparing emotional expressions across large text pieces.
- Applications: The resulting distributions can support applications such as clustering literary texts and detecting emotion-word spikes near target mentions.The paper also visualizes emotion percentages and differences between Shakespeare’s Hamlet and As you like it.
4 Visualizations of Emotions
The paper uses comparative bars, relative-salience word clouds, and timelines to visualize emotion distributions, sources, and changes across literary texts.
- Comparative distributions: Hamlet contains more fear, sadness, disgust, and anger, while As You Like It contains more joy, trust, and anticipation.
- Comparative distributions: Relative-salience word clouds reveal which trust and sadness words distinguish Hamlet from As You Like It.The comparison uses word frequencies normalized by each text’s total word count.
- Emotion flow: Emotion timelines track the flow of joy, trust, and fear through As You Like It, Hamlet, and Frankenstein.
- Emotion flow: The visualizations portray the novels as progressively darker from comedy to tragedy to horror, with Frankenstein darkest in its final chapters.
5 Emotion Word Density
Emotion word density measures expected emotion-word counts per fixed number of words and supports organizing and comparing fairy tales through interactive visualizations.
- Emotion word density is the number of emotion words expected in every X words; this paper reports densities for X = 10,000.
- The Brothers Grimm fairy tales are arranged by increasing negative word density across 192 stories.The plot also displays joy and sadness densities.
- Selecting two tales reveals their differences in emotion percentages through a comparison bar graph.Cinderella and Godfather Death are used as an example, with a corresponding joy-word relative-salience cloud.
6 Emotions Associated with Targets
The paper estimates emotions associated with entities by analyzing nearby words in Google Books 5-grams, revealing historical patterns in fear associations.
- The system identifies 5-grams containing a target, extracts nearby emotion words, and calculates their percentages.
- Google Books 5-gram data covers books published from 1600 to 2009; the analysis excludes books before 1800 because coverage is less comprehensive.
- Fear-word percentages rise around World War I for Germany, America, and China, while China also shows a spike around 1900.
- India shows no World War I spike but has a 1940s increase associated with its independence struggle and growing involvement in World War II.
- The analysis also compares joy and anger word percentages near occurrences of woman and man in books.
7 Emotion Words in Novels vs. Fairy Tales
The study compares emotion and polarity word densities in novels and fairy tales, finding broader variation and several significant density differences in fairy tales.
- Corpora and measures: The study uses the Corpus of English Novels and Fairy Tale Corpus to compare emotion and polarity word densities.The novels corpus contains 292 novels, while the fairy-tale corpus contains 453 stories.
- Emotion densities: Fairy tales have significantly higher anticipation, disgust, joy, and surprise densities than novels, but significantly lower trust density.Anger and sadness densities do not differ significantly between the corpora; the reported differences have p < 0.001.
- Emotion densities: Fairy tales show significantly greater standard deviations than novels for all eight emotions, producing more texts with both high and low emotion densities.This wider spread is a central difference between the two corpora.
- Polarity densities: Fairy tales have significantly fewer negative terms and significantly more positive words than novels, with p < 0.001.For comparison, the novels corpus averages about 1670 negative words per 10,000 words.
- Distribution analysis: Histograms show that fairy-tale emotion densities are not bimodal; instead, more fairy tales than novels have extreme emotion densities.The distribution supports a broad spread rather than two distinct clusters.
8 Conclusions and Future Work
The paper presents an emotion-analysis and visualization framework for literary texts, including emotion word density and corpus-scale entity associations. It also identifies fairy tales as having a wider distribution of emotion word densities than novels and situates the work within an affect-based search interface.
- Conclusions: The paper combines an emotion analyzer, visualizations, emotion word density, and Google Books Corpus analysis to study literary texts and organize them for search.The framework addresses individual books and very large collections.
- Conclusions: Fairy tales have a much wider distribution of emotion word densities than novels.This comparison is presented as a contribution using the emotion lexicon.
- Future work: The broader project aims to provide an affect-based Project Gutenberg interface supporting emotion-oriented plots and searches for snippets with strong emotion word densities.The intended interface takes a search query and returns relevant plots or snippets from multiple texts.