Source-linked AI summary
Measuring the happiness of large-scale written expression: Songs, Blogs, and Presidents
Peter Sheridan Dodds, Christopher M. Danforth
TL;DR
Population-level emotional states are important to measure, but existing approaches can be invasive, unreliable, or small-scale. The paper uses human word-valence ratings to compute continuous happiness scores for large text corpora, finding temporal and demographic patterns across songs, blogs, and presidential addresses.
Problem
Existing happiness measures rely heavily on self-reports that can be invasive, unreliable, and limited to small samples, motivating a transparent population-level alternative.
Method
The method computes a frequency-weighted average of human-assessed ANEW psychological valences for words appearing in large-scale texts.
Results
Song-lyric happiness declined from 1961 to 2007, with the decline strongest before 1985 and leveling after 1995, while blog valence varied with age and latitude.
Takeaways & Limitations
Continuous word-valence scores provide a transparent, scalable way to quantify emotional content across large written corpora.
Abstract
from arXiv · showhide
The importance of quantifying the nature and intensity of emotional states at the level of populations is evident: we would like to know how, when, and why individuals feel as they do if we wish, for example, to better construct public policy, build more successful organizations, and, from a scientific perspective, more fully understand economic and social phenomena. Here, by incorporating direct human assessment of words, we quantify happiness levels on a continuous scale for a diverse set of large-scale texts: song titles and lyrics, weblogs, and State of the Union addresses. Our method is transparent, improvable, capable of rapidly processing Web-scale texts, and moves beyond approaches based on coarse categorization. Among a number of observations, we find that the happiness of song lyrics trends downward from the 1960's to the mid 1990's while remaining stable within genres, and that the happiness of blogs has steadily increased from 2005 to 2009, exhibiting a striking rise and fall with blogger age and distance from the equator.
I. INTRODUCTION
The paper develops a transparent, scalable way to estimate happiness in large-scale written expression using human ratings of individual words. It addresses limitations of self-reported and coarse-category approaches while emphasizing that the measure is intended for robust analysis of large texts.
- Self-reported happiness measures can be invasive, memory- and self-perception-dependent, prone to misreporting, and limited to small samples.
- The method uses human evaluations of words within texts to generate an overall emotional score, offering a non-reactive population-level measure.
- ANEW participants rated 1034 emotionally meaningful words for psychological valence on a 1–9 scale with half-integer increments.
- The text valence score is a frequency-weighted average of ANEW word valences, using each word’s occurrence count in the text.
- Because word combinations are not interpreted, the method is suitable only for large-scale texts, where aggregation supports robust statistical analysis.
- The approach quantifies happiness continuously rather than assigning texts to broad positive, neutral, or negative categories, and can be refined by future human-response studies.
II. DESCRIPTION OF LARGE-SCALE TEXTS STUDIED
The study examines song lyrics, song titles, blog sentences, and State of the Union addresses using large-scale text sources and ANEW-word valence distributions. These corpora differ in coverage, authorship, frequent words, and the distribution of emotional valence.
- Corpus coverage: 3.5% to 9.2%: ANEW words comprise these percentages of the four corpora, from State of the Union addresses to song titles.The datasets also differ in total word counts and numbers of individual artists, blogs, and presidents.
- Frequent words: “Love” dominates song lyrics and titles, while “people” and “life” characterize blogs and “world” and “war” appear disproportionately in State of the Union addresses.These top-five ANEW words are expressed as frequencies among ANEW words in each corpus.
- Valence distributions: Song lyrics are weighted toward high-valence words, whereas blogs contain more low-valence words and have a bimodal distribution with a mode at 8–9.The comparison uses normalized abundances of ANEW words binned by average valence, with Michael Jackson’s lyrics included as a reference example.
III. RESULTS
Song lyrics show a robust decline in average valence from 1961 to 2007, driven by changing word frequencies and the emergence of lower-valence genres. Blog valence increased from 2005 to 2009 and varied systematically with dates, age, latitude, weekdays, and country.
- Songs: Song-lyric valence declined from 1961 to 2007, with the steepest decrease before about 1985 and an apparent leveling after 1995.The downward trend persisted across 100 random 750-word ANEW subsets, supporting the stability of relative changes.
- Songs: Individual music genres remained relatively stable in valence over time, while newer genres such as metal and punk occupied lower-valence emotional niches.The overall decline therefore occurred across the evolving genre composition rather than within most particular genres.
- Blogs: Blog-sentence valence rose from around 5.75 to above 6.0 over the examined period, with within-year increases often concentrated late in the year.The 2008 series included a midyear dip followed by a late-year peak.
- Blogs: Blog valence followed a convex age pattern, rising from lows of 5.55 among 14-year-olds toward about 6.0 at ages 45–60 before declining.The comparison of bloggers turning 14 and those aged 45–60 reports average valences of 5.55 and 5.98, respectively.
- Blogs: Blog valence also varied with events and context: it ranged from 5.71 near the equator to 5.83 at mid-latitudes, peaked on Sundays, and was highest in the United States among four represented countries.The country averages were United States 5.83, Canada 5.78, United Kingdom 5.77, and Australia 5.74.
IV. CONCLUDING REMARKS
The paper identifies expanding online records of social interactions and personal experiences as opportunities for further scientific investigation. It also outlines several ways to improve the method, including richer word assessments, participant demographics, alternative evaluation approaches, and phrase-level measurements.
- Growing online records of social interactions and personal experiences will provide richer datasets for scientific investigations.The authors propose examining emotional dynamics in online interactive contexts, including possible emotional contagion.
- The method could be improved by collecting emotional-content estimates for a more extensive set of individual words.
- Participant demographics could be incorporated into the word-assessment instrument to make it more sophisticated.
- Alternative approaches beyond ANEW-style semantic differentials could be explored.
- Assessing common word groups and phrases could clarify the micro-macro connection between words and sentences and differences across age groups and cultures.