Affective biases in English are bi-dimensional

Abstract

A long-standing observation about the interface between emotion and language is that positive words are used more frequently than negative ones, leading to the Pollyanna hypothesis which alleges a predominantly optimistic outlook in humans. This paper uses the largest available collection of affective ratings as well as insights from linguistics to revisit the Pollyanna hypothesis as it relates to two dimensions of emotion: valence (pleasantness) and arousal (intensity). We identified systematic patterns in the distribution of words over a bi-dimensional affective space, which (1) run counter to and supersede most prior accounts, and (2) differ drastically between word types (unique, distinct words in the lexicon) and word tokens (number of occurrences of available words in the lexicon). We argue for two factors that shape affect in language and society: a pro-social benevolent communication strategy with its emphasis on useful and dangerous phenomena, and the structure of human subjective perception of affect.

Keywords:

• Valence
• Arousal
• Lexicon
• Cognitive bias
• Subjective experience

This work was supported by an SSHRC Joseph-Armand Bombardier CGS Scholarship to the first author and the SSHRC Insight Development [grant number 430-2012-0488], the NSERC Discovery [grant number 402395-2012] and the NIH R01 HD 073288 (PI Julie A. Van Dyke) to the second author.

Notes

¹ Within a given corpus of texts, a word's type frequency will always be 1. But if we group words into affective categories then we can compare type frequencies by category. For example, a corpus will have a certain number of positive word types and a certain number of negative word types. An individual word's frequency reported by a corpus refers to word tokens, i.e., the number of instances of that word that appear within the entire corpus.

² We are indebted to an anonymous reviewer for raising this point.

³ Kuppens et al. (2013) point out that the U- or a V-shaped relationship observed in ratings aggregated over hundreds or thousands individuals co-exists with very large individual variability in the relationship between arousal and valence. We are not able to test this between-levels difference as individual frequency distributions are not available to us.