“Nymph Piss and Gravy Orgies: Local and Global Contrast Effects in Relational Humor”, Cynthia S. Q. Siew, Tomas Engelthaler, Thomas T. Hills2022-04-11 (, , )⁠:

[cf. Reilly et al 2020 on profanity] How does the relation between two words create humor? In this article, we investigated the effect of global and local contrast on the humor of word pairs.

We capitalized on the existence of psycholinguistic lexical norms by examining violations of expectations set up by typical patterns of English usage (global contrast) and within the local context of the words within the word pairs (local contrast). Global contrast was operationalized as lexical-semantic norms for single-words and local contrast was operationalized as the orthographic, phonological, and semantic distance between the two words in the pair.

Through crowd-sourced (Study 1) and best–worst (Study 2) ratings of the humor of a large set of word pairs (ie. compounds), we find:

Table 2.1: Top 10 Least Humorous Word Pairs From Study 2.
Least humorous Predicted probability of humor
sell bargain 0.288
conserve health 0.289
power influence 0.291
will stay 0.298
schedule year 0.303
insult nickname 0.322
life friend 0.323
trouble mention 0.324
workman call 0.326
large small 0.327
Table 2.2: Top 10 Most Humorous Word Pairs From Study 2.
Most humorous Predicted probability of humor
polka hooker 0.765
playboy parrot 0.755
penis weasel 0.745
turnip tramp 0.714
funk fungus 0.714
spam scrotum 0.709
gnome bone 0.697
stripper hippo 0.694
rowdy bowels 0.693
pansy panties 0.693

evidence of both global and local contrast on compound-word humor. Specifically, we find that humor arises when there is a violation of expectations at the local level, between the individual words that make up the word pair, even after accounting for violations at the global level relative to the entire language. Semantic variables (arousal, dominance, and concreteness) were stronger predictors of word pair humor whereas form-related variables (number of letters, phonemes, and letter frequency) were stronger predictors of single-word humor.

Moreover, we also find that semantic dissimilarity increases humor, by defusing the impact of low-valence words—making them seem more amusing—and by enhancing the incongruence of highly imageable pairs of concrete words.

[Keywords: compound-word humor, semantic similarity, phonological distance]

…For example, which is funnier, the word porridge or the word oatmeal? Most people agree that ‘porridge’ is funnier than ‘oatmeal’. This may at first glance appear to violate a relational theory of humor because it is not obvious what the context is for a word on its own. However, the data from Engelthaler & Hills2018 suggest that the violation may be as simple as word frequency. Lower frequency words tend to be rated as more humorous than higher frequency words; inverse frequency is the strongest predictor of single word humor. Westbury & Hollis2019 go on to show that low probability orthographic or phonological structure are also well correlated with humor of individual words, further suggesting that single word humor is the outcome of a cognitive process for entropy detection.

The natural extension of single word humor is to ask if these results scale up to multi-word humor. In this article, we address this question by building upon the prior work of Engelthaler & Hills2018 and Westbury & Hollis2019, making a simple alteration of their prior research on single words by adding a second word. Now instead of facing our participants with the task of rating individual words, like cage (that is not particularly funny on its own) or cabbage (only mildly funnier), our participants are faced with rating the humor of cabbage cage, which is arguably funnier than either word alone. But why?

…Because the number of possible word pairs that could be generated from even a limited set of words (ie. using all 4,997 words from the Engelthaler & Hills2018 single-word humor norms would result in 4,9972 = 25 million pairs) was very large, we deliberately adopted an approach that crowdsourced humor ratings from volunteers who viewed randomly generated pairs of words on a web application.

[Probably language models like GPT-3 could screen for humor.]