Human language reveals a universal positivity bias

Results and Discussion

In Fig. 1, we show distributions of the average happiness scores for all 24 corpora, leading to our most general observation of a clear positivity bias in natural language. We indicate the above-neutral part of each distribution with yellow and the below-neutral part with blue, and we order the distributions moving upward by increasing median (vertical red line). For all corpora, the median clearly exceeds the neutral score of 5. The background gray lines connect deciles for each distribution. In SI Appendix, Fig. S1, we provide the same distributions ordered instead by increasing variance.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 1.

Distributions of perceived average word happiness havg for 24 corpora in 10 languages. The histograms represent the 5,000 most commonly used words in each corpora (details are provided in SI Appendix), and native speakers scored words on a 1–9 double-Likert scale, with 1 being extremely negative, 5 being neutral, and 9 being extremely positive. Yellow indicates positivity (havg>5), and blue indicates negativity (havg<5). Distributions are ordered by increasing median (red vertical line). The background gray lines connect deciles of adjacent distributions. The same distributions arranged according to increasing variance are shown in SI Appendix, Fig. S1.

As is evident from the ordering in Fig. 1 and SI Appendix, Fig. S1, although a positivity bias is the universal rule, there are minor differences between the happiness distributions of languages. For example, Latin American-evaluated corpora (Mexican Spanish and Brazilian Portuguese) exhibit relatively high medians and, to a lesser degree, higher variances. For other languages, we see that those languages with multiple corpora have more variable medians, and specific corpora are not ordered by median in the same way across languages (e.g., Google Books has a lower median than Twitter for Russian, but the reverse is true for German and English). In terms of emotional variance, all four English corpora are among the highest, whereas Chinese and Russian Google Books seem especially constrained.

We now examine how individual words themselves vary in their average happiness score between languages. Owing to the scale of our corpora, we were compelled to use an online service, choosing Google Translate. For each of the 45 language pairs, we translated isolated words from one language to the other and then back. We then found all word pairs that (i) were translationally stable, meaning the forward and back translation returns the original word, and (ii) appeared in our corpora for each language.

We provide the resulting comparison between languages at the level of individual words in Fig. 2. We use the mean of each language’s word happiness distribution derived from its merged corpora to generate a rough overall ordering, acknowledging that frequency of use is no longer meaningful, and moreover is not relevant, because we are now investigating the properties of individual words. Each cell shows a heat map comparison with word density increasing as shading moves from gray to white. The background colors reflect the ordering of each pair of languages: yellow if the row language had a higher average happiness than the column language and blue for the reverse. Also, in each cell, we display the number of translation-stable words between language pairs, N, along with the difference in average word happiness, Δ, where each word is equally weighted.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 2.

Scatter plots of average happiness for words measured in different languages. We order languages from relatively most positive (Spanish) to relatively least positive (Chinese); a yellow background indicates the row language is more positive than the column language, and a blue background indicates the converse. The overall plot matrix is symmetrical about the leading diagonal, with the redundancy allowing for easier comparison between languages. In each scatter plot, the key gives the number of translation-stable words for each language pair, N; the average difference in translation-stable word happiness between the row language and the column language, Δ; and the Pearson correlation coefficient for the regression, r. All P values are less than 10−118 for the Pearson correlation coefficient and less than 10−82 for the Spearman correlation coefficient.

A linear relationship is clear for each language–language comparison, and is supported by Pearson’s correlation coefficient r being in the range 0.73–0.89 (P <10⁻¹¹⁸ across all pairs; Fig. 2 and SI Appendix, Tables S3–S5). Overall, this strong agreement between languages suggests that approximate estimates of word happiness for unscored languages could be generated with no expense from our existing dataset. Some words will, of course, translate unsatisfactorily, with the dominant meaning changing between languages. For example “lying” in English, most readily interpreted as speaking falsehoods by our participants, translates to “acostado” in Spanish, meaning recumbent. Nevertheless, happiness scores obtained by translation will be serviceable for purposes where the effects of many different words are incorporated.

Stepping back from examining interlanguage robustness, we return to a more detailed exploration of the rich structure of each corpus’s happiness distribution. In Fig. 3, we show how average word happiness havg is largely independent of word use frequency for four example corpora. We first plot use frequency rank r of the 5,000 most frequently used words as a function of their average happiness score, havg (background dots), along with some example evenly spaced words. We note that words at the extremes of the happiness scale are ones evaluators agreed upon strongly, whereas words near neutral range from being clearly neutral [e.g., havg (“the”) = 4.98] to contentious with high SD (17). We then compute deciles for contiguous sets of 500 words, sliding this window through rank r. These deciles form the vertical strands. We overlay randomly chosen, equally spaced example words to give a sense of each corpus’s emotional texture.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 3.

Examples of how word happiness varies little with use frequency. (A–D) Above each plot is a histogram of average happiness havg for the 5,000 most frequently used words in the given corpus, matching Fig. 1. Each point locates a word by its usage rank r and average happiness havg, and we show some regularly spaced example words. The descending gray curves of these jellyfish plots indicate deciles for windows of 500 words of contiguous use rank, showing that the overall histogram’s form is roughly maintained at all scales. The “kkkkkk…” words represent laughter in Brazilian Portuguese, in the manner of “hahaha…,” and an English translation is provided in SI Appendix, Fig. S2.

We chose the four example corpora shown in Fig. 3 to be disparate in nature, covering diverse languages (French, Egyptian Arabic, Brazilian Portuguese, and Chinese), regions of the world (Europe, the Middle East, South America, and Asia), and texts [Twitter, movies and television, the web (15), and books (14)]. The remaining 20 corpora yield similar plots (SI Appendix), and all corpora also exhibit an approximate self-similarity in SD for word happiness.

Across all corpora, we observe visually that the deciles tend to stay fixed or move slightly toward the negative, with some expected fragility at the 10% and 90% levels (due to the distributions’ tails), indicating that the overall happiness distribution of each corpus approximately holds independent of word use. In Fig. 3, for example, we see that both the Brazilian Portuguese and French examples show a small shift to the negative for increasingly rare words, whereas there is no visually clear trend for the Arabic and Chinese cases. Fitting havg=αr+β typically returns α on the order of −1 ×10−5 suggesting havg decreases 0.1 per 10,000 words. For SDs of happiness scores, we find a similarly weak drift toward higher values for increasingly rare words (correlations and linear fits for havg and hstd as a function of word rank r for all corpora are provided in SI Appendix, Tables S6 and S7). We thus find that, to first order, not just the positivity bias but the happiness distribution itself applies for common words and rare words alike, revealing an unexpected addition to the many well-known scalings found in natural language, as famously exemplified by Zipf’s law (22).

In constructing language-based instruments for measuring expressed happiness, such as our hedonometer (18), this frequency independence allows for a way to “increase the gain” in a fashion resembling standard physical instruments. Moreover, we have earlier demonstrated the robustness of our hedonometer for the English language, showing, for example, that measurements derived from Twitter correlate strongly with Gallup well-being polls and related indices at the state and city level for the United States (19).

Here, we provide an illustrative use of our hedonometer in the realm of literature, inspired by Vonnegut’s shapes of stories (23, 24). In Fig. 4, we show “happiness time series” for three famous works of literature, evaluated in their original languages of English, Russian, and French, respectively: Melville’s Moby Dick (www.gutenberg.org), Dostoyevsky’s Crime and Punishment (25), and Dumas’ The Count of Monte Cristo (www.gutenberg.org). We slide a 10,000-word window through each work, computing the average happiness using a “lens” for the hedonometer in the following manner. We capitalize on our instrument’s tunability to obtain a strong signal by excluding all words for which 3<havg<7 (i.e., we keep words residing in the tails of each distribution) (18). Denoting a given lens by its corresponding set of allowed words L, we estimate the happiness score of any text T as havg(T)=∑w∈Lfwhavg(w)/∑w∈Lfw, where fw is the frequency of word w in T (16).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 4.

Emotional time series for three famous 19th century works of literature: Melville’s Moby Dick (Top), Dostoyevsky’s Crime and Punishment (Middle), and Dumas’ The Count of Monte Cristo (Bottom). Each point represents the language-specific happiness score for a window of 10,000 words (converted to lowercase), with the window translated throughout the work. The overlaid word shifts (A–I) show example comparisons between different sections of each work. Word shifts indicate which words contribute the most toward and against the change in average happiness between two texts (SI Appendix, pp. S9–S10). Although a robust instrument in general, we acknowledge the hedonometer’s potential failure for individual words due to both language evolution and words possessing more than one meaning. For Moby Dick, we excluded “cried” and “cry” (to speak loudly rather than weep) and “Coffin” (surname, still common in Nantucket). Such alterations, which can be done on a case-by-case basis, do not noticeably change the overall happiness curves, although leaving the word shifts more informative. We provide online, interactive visualizations of the emotional trajectories of over 10,000 books at hedonometer.org/books.html.

The three resulting happiness time series provide interesting detailed views of each work’s narrative trajectory revealing numerous peaks and troughs throughout, at times clearly dropping below neutral. Both Moby Dick and Crime and Punishment end on low notes, whereas The Count of Monte Cristo culminates with a rise in positivity, accurately reflecting the finishing arcs of all three. The “word shifts” overlaying the time series compare two distinct regions of each work, showing how changes in word abundances lead to overall shifts in average happiness. Such word shifts are essential tests of any sentiment measurement, and are made possible by the linear form of our instrument (16, 18) [a full explanation is provided in SI Appendix, pp. S9–S10]. As one example, the third word shift for Moby Dick shows why the average happiness of the last 10% of the book is well below the average happiness of the first 25%. The major contribution is an increase in relatively negative words, including “missing,” “shot,” “poor,” “die,” and “evil.”

By adjusting the lens, many other related time series can be formed, such as those produced by focusing on only positive or negative words. Emotional variance as a function of text position can also be readily extracted. At hedonometer.org/books.html, we provide online, interactive emotional trajectories for over 10,000 works of literature where different lenses and regions of comparisons may be easily explored. Beyond this example tool we have created here for the digital humanities and our hedonometer for measuring population well-being, the datasets we have generated for the present study may be useful in creating a great variety of language-based instruments for assessing emotional expression.

Overall, our major scientific finding is that when experienced in isolation and weighted properly according to use, words, which are the atoms of human language, present an emotional spectrum with a universal, self-similar positive bias. We emphasize that this apparent linguistic encoding of our social nature is a system-level property, and in no way asserts all natural texts will skew positive (as exemplified by certain passages of the three works in Fig. 4) or diminishes the salience of negative states (26). Going forward, our word happiness assessments should be periodically repeated and carried out for new languages, tested on different demographics, and expanded to phrases both for the improvement of hedonometric instruments and to chart the dynamics of our collective social self.