Source-linked AI summary
Positive words carry less information than negative words
David Garcia, Antonios Garas, Frank Schweitzer
TL;DR
The paper examines whether emotional valence shapes written word use and information content, a relevant baseline question for sentiment analysis. It relates valence lexica to word and sequence frequencies, finding that positive words are used more often while negative words carry more information.
Problem
The study asks whether written emotional expression has a positive bias, a baseline question for sentiment analyses that often assume neutral expression.
Source: §1 · pp. 1–2Method
The authors relate valence ratings from three established language lexica to web word frequencies and estimate information from word and N-gram frequencies.
Source: §1 · pp. 1–2, §2.2 · pp. 5, 8Results
Positive words are used more often across the three studied languages, while negative words carry more self-information even with contexts up to four-word sequences.
Source: §3 · p. 11Takeaways & Limitations
The authors suggest positive emotional expression may reinforce social links, while negative expressions convey highly informative and relevant events.
Source: §3 · p. 12The analysis leaves syntax's role in emotional expression beyond scope and excludes function words, whose presence can modify sentence-level emotional content.
Source: §3 · pp. 11–12
Abstract
from arXiv · showhide
We show that the frequency of word use is not only determined by the word length \cite{Zipf1935} and the average information content \cite{Piantadosi2011}, but also by its emotional content. We have analyzed three established lexica of affective word usage in English, German, and Spanish, to verify that these lexica have a neutral, unbiased, emotional content. Taking into account the frequency of word usage, we find that words with a positive emotional content are more frequently used. This lends support to Pollyanna hypothesis \cite{Boucher1969} that there should be a positive bias in human expression. We also find that negative words contain more information than positive words, as the informativeness of a word increases uniformly with its valence decrease. Our findings support earlier conjectures about (i) the relation between word frequency and information content, and (ii) the impact of positive emotions on communication and social links.
1 Introduction
The paper asks whether emotional content shapes word use and information alongside established word-length and information relationships. Using affective lexica and word-frequency data, it finds a positive bias in usage and links emotional valence to informativeness.
1 Introduction
- The study combines word-valence estimates with usage frequencies from the indexable web to quantify the baseline of written emotional expression.It covers English, German, and Spanish, which together span more than 67.7% of websites.Source: §1 · p. 2
- Positive words appear more frequently, while negative words carry more information, supporting a positive bias in language use.The authors relate this pattern to the Pollyanna hypothesis.Source: §1 · pp. 1–2
- The analysis extends information-content work by examining emotional valence and word context, finding that positive words carry less information than negative ones.The introduction frames emotional content as affecting word frequency and informativeness.Source: §1 · p. 2
- The authors favor established, controlled valence lexica over an earlier uncontrolled study, whose positive bias could reflect response and wording biases.The lexica contain 1,000 to 3,000 words and were produced in three independent controlled setups.Source: §1 · pp. 2–3
2 Results
Across English, German, and Spanish lexica, frequency-weighted usage is more positive than the lexica’s near-neutral valence distributions. Valence also decreases with word information, with additional analyses separating this relationship from word length and self-information.
2.1 Frequency of emotional words
- 0.257±0.032 for English, 0.167 ± 0.017 for German, and 0.287 ± 0.035 for Spanish were the estimated 95% confidence intervals for frequency-weighted mean-valence differences.The frequency-weighted distributions shift toward positive valence, making the usage-based emotional reference point about 0.3 rather than zero.Source: §2.1 · p. 5, Figure 2 · p. 6
2.2 Relation between information and valence
- For all three languages, valence correlates more strongly in absolute value with self-information than with plain word frequency.This comparison indicates that self-information provides more knowledge about word valence than frequency alone.Source: §2.2 · pp. 5–6
- The valence–self-information relationship is similar but slightly weaker when estimated from book frequencies, supporting robustness across online and traditional written text.The control analysis used English, German, and Spanish book-frequency datasets.Source: §2.2 · p. 7
- Across the three languages, valence decreases as word self-information and context-based information content increase, so positive words carry less information than negative words.The pattern is reported for self-information and for context sizes up to four, although the four-gram trend is less clear for German.Source: §2.2 · p. 8, Figure 3 · p. 9
2.3 Additional analysis of valence, length and self-information
- Word length increases with self-information, but valence–self-information correlations remain within their original 95% confidence intervals after controlling for length.The reported correlations between valence and length are very low or not significant, supporting valence as an additional semantic dimension rather than a proxy for word length.Source: §2.3 · p. 10, Table 2 · p. 10
- After controlling for self-information, most context-size correlations between valence and information content remain significant and negative, except for English I2.The authors suggest the weaker two-word-context correlations may relate to constructions such as negations, articles before nouns, or epithets; I3 and I4 add information about valence beyond self-information.Source: §2.3 · pp. 10–11, Table 3 · p. 11
3 Discussion
Across three languages and multiple corpora, positive words are used more often, while rarer negative words carry more information. The discussion relates this pattern to communication, social links, and the scope of word-level analysis.
- Positive words are used more often across three languages and corpora, while negative words are rarer and carry more self-information even across four-word sequences, independent of length.The authors report that the relation holds for traditional written-text corpora and Google data.Source: §3 · p. 11
- Syntax was not explored directly; function words were excluded despite their capacity to alter sentence-level emotional meaning, and sentence structure also correlated with emotional content.The authors suggest future work incorporate valence scores for word sequences and examine syntactical rules.Source: §3 · pp. 11–12
- The authors suggest communication's information-transfer function may favor positive content, while negative expressions transmit highly informative and relevant events.They offer the formation of social links as a possible explanation for the positive bias.Source: §3 · p. 12
- They connect frequent positive expression to positivity offset and greater informativeness of negative expression to negativity bias, presenting these as explanatory accounts.The discussion also suggests people may avoid negative topics and bring up positive ones because it feels better to talk about pleasant things.Source: §3 · p. 12
- The estimated emotional baseline may inform online-text analyses and anomaly detection; the authors propose fake human-like texts may fail to reproduce this balance.The discussion says automatic tools and further analyses will need to account for the positive emotional charge of written expression.Source: §3 · pp. 12–13