Source-linked AI summary
Positivity of the English language
Isabel M. Kloumann, Christopher M. Danforth, Kameron Decker Harris, Catherine A. Bliss, Peter Sheridan Dodds
TL;DR
Prior findings on linguistic positivity are mixed, leaving the emotional spectrum and overall bias of natural language open questions. The paper rates frequency-selected English words across four corpora and finds positive words predominate, with positivity distributions broadly consistent across frequency ranks.
Problem
Prior evidence on English-language positivity is mixed, with positive-bias findings from small word samples and a negative bias in experimental emotional-word elicitation.
Source: Introduction · p. 1Method
The study selects words by frequency from four large English text corpora and measures their average human-rated happiness.
Source: Introduction · p. 1, Results and Discussion · p. 2Results
Positive words comprise 72.00% (TW), 78.80% (GB), 78.38% (NYT), and 64.14% (ML), and distribution forms show internal consistency across usage frequency.
Source: Results and Discussion · p. 2Takeaways & Limitations
The four corpora show a clear positive bias, while their word-positivity distributions remain broadly consistent across usage-frequency ranks.
Source: Results and Discussion · p. 2
Abstract
from arXiv · showhide
Over the last million years, human language has emerged and evolved as a fundamental instrument of social communication and semiotic representation. People use language in part to convey emotional information, leading to the central and contingent questions: (1) What is the emotional spectrum of natural language? and (2) Are natural languages neutrally, positively, or negatively biased? Here, we report that the human-perceived positivity of over 10,000 of the most frequently used English words exhibits a clear positive bias. More deeply, we characterize and quantify distributions of word positivity for four large and distinct corpora, demonstrating that their form is broadly invariant with respect to frequency of word use.
Introduction
The introduction frames language as a source of evidence about human social and emotional life, while noting uncertainty about human cooperation and mixed findings on linguistic positivity. The study tests English word positivity using frequency-selected words across four corpora.
Introduction
- Traditional economic theory treats people as rationally selfish, while behavioral findings and philosophical traditions offer competing accounts of human cooperation and pro-social behavior.The introduction describes uncertainty about how cooperative individuals or groups are, or should be.Source: Introduction · p. 1
- Previous studies offer mixed evidence on linguistic positivity: small English-word samples suggest a positive bias, while emotional-word elicitation finds a strong negative bias.The positive-bias findings are framed as the Pollyanna Hypothesis and Linguistic Positivity Bias.Source: Introduction · p. 1
- The study selects words solely by frequency and measures happiness, or psychological valence, across Twitter, Google Books, The New York Times, and music lyrics.The corpora span broadcast media, opinion, literature, songs, and public social interactions, with differing grammatical and orthographic correctness.Source: Introduction · p. 1
- The authors combine the 5,000 most frequent words from each corpus into 10,222 unique words and collect 50 independent happiness ratings per word on a 1–9 scale.Participants rate each word in isolation; 1 is least happy, 5 neutral, and 9 most happy.Source: Introduction · p. 1
- For the labMT 1.0 data set, the highest and lowest average happiness scores were 8.50 for laughter and 1.30 for terrorist; neutral words averaged near 5.The averages were 4.98 for the and 5.02 for it.Source: Introduction · p. 1
Results and Discussion
All four corpora show a positively skewed distribution of word happiness, with positive words more abundant than negative ones. Positivity distributions are broadly consistent across frequency bands, though frequency-related patterns in average happiness and rating variation differ by corpus.
Consistency across frequency bands
- Of 40 pairwise comparisons between frequency bands, 29 show statistically significant distribution matches (p > 10−2), with neighboring bands often likely to match.The NYT, ML, and GB show the most internal agreement.Source: Results and Discussion · p. 2
Corpus-specific patterns
- The most frequent 5000 words overlap across corpora, but numerous words occur in only one corpus and the frequency rankings differ strongly.For example, ‘rainbows’ and ‘kissing’ appear in the Music Lyrics top 5000, while ‘punishment’ appears in the Google Books top 5000.Source: Results and Discussion · pp. 2–3
Overall positivity
- Positive words make up 72.00% (TW), 78.80% (GB), 78.38% (NYT), and 64.14% (ML) of each corpus’s most frequent 5000 words.The four distributions are unimodal and strongly positively skewed; the neutral score falls below the median.Source: Results and Discussion · p. 2, Figure 1 · p. 3
Average happiness by frequency
- In the four corpora, average-happiness slopes are negative but extremely small, ranging from -3.04×10−5 (GB) to -7.78×10−5 (TW).All four corpora also have weak negative correlations between average happiness and usage-frequency rank; GB’s correlation is not statistically significant (p=0.35).Source: Results and Discussion · p. 3
- Average happiness is nearly stable across frequency ranks in GB and NYT, while TW and ML show a slight trend toward lower happiness for rarer words.The reported pattern is small enough to be likely indiscernible to an individual familiar with these corpora.Source: Results and Discussion · pp. 4–5, Figure 3 · p. 5
Variation in ratings
- Words with high rating standard deviations include terms related to profanities, alcohol and tobacco, religion, politics, sex, marriage, fast food, climate, and cultural phenomena.The authors suggest diverse ratings may reflect raters’ preferences or cultural and demographic variation.Source: Results and Discussion · pp. 5–6
- Rating standard deviations tend to rise with frequency rank in GB and ML, while TW shows no evidence of growth and NYT has no obvious linear variation.The increase is gradual in GB; relatively rarer words have higher standard deviations there than more common words.Source: Results and Discussion · pp. 6–7
Corpus-specific patterns
- The four corpora’s slight frequency-related deviations differ by measure: TW and ML show negative average-happiness correlations, while GB and ML show positive standard-deviation correlations.The authors characterize these deviations as highly constrained and corpus specific.Source: Results and Discussion · p. 7
Concluding remarks
Positive words strongly outnumber negative words overall, while any tendency for high-frequency words to be more positive is very limited and corpus-specific.
- Positive words strongly outnumber negative words overall, while any tendency for high-frequency words to be more positive is very limited and corpus-specific.Source: Concluding remarks · p. 8
Supplementary Information
The supplementary material includes statistical comparisons of happiness distributions, plots of example words and standard deviations, and ranked word tables.
- Kolmogorov-Smirnov test p-values compare happiness distributions across pairs of frequency ranges within each corpus.Lighter colors indicate a greater probability that the samples could come from the same distribution; grayscale represents log10(p-value).Source: Figure S1 · p. 11
- Supplementary plots show example words by frequency rank and average happiness for Twitter, Google Books, and Music Lyrics.Source: Figure S2 · p. 12, Figure S3 · p. 13, Figure S4 · p. 14
- Supplementary plots show overall standard-deviation distributions and example words by frequency rank for Twitter, New York Times, and Music Lyrics.Source: Figure S5 · p. 15, Figure S6 · p. 16, Figure S7 · p. 17
- Supplementary tables list the 50 most positive words, 50 most negative words, and 50 words with the highest standard deviation of happiness estimates.Source: Table S1 · p. 19, Table S2 · p. 20, Table S3 · p. 21