What’s in a pun? Assessing the relationship between phonological distance and perceived funniness of punning jokes

1 Introduction

Punning is a form of humorous wordplay based on semantic ambiguity between two words that are phonologically similar or identical, in a syntactic context in which both meanings are more or less acceptable (Hempelmann and Miller 2017). Humor in punning jokes relies strongly on phonological features of both the pun word, which is explicitly expressed in the joke, and the target word, whose meaning is invoked implicitly. Several attempts have been made to quantify those features and investigate possible correlations with the comprehensibility of a punning joke (for an overview of which see Hempelmann and Miller 2017). However, to date there has been no empirical investigation of the relationship of phonological features between puns and their targets and the perceived funniness of the joke that contains them. The study described in this article investigates this question, in particular through analyses of correlations between funniness ratings^[1] of punning jokes obtained in a crowdsourcing study and various phonological distance measures of the respective pun–target pairs.

From previous analyses of puns (e.g., Lagerquist 1980), we expect punning jokes with lower phonological distance between the pun and target word to be associated with higher funniness ratings. More simply put, the more the pun and its target sound alike, the funnier the punning joke should be perceived as. Consequently, we would expect homophonic puns – those in which the pun and target sound exactly the same – to be perceived as funnier than heterophonic ones, in which the pun and target are further apart in sound. The statistical analyses we present in this paper provide empirical evidence for this tendency. As we discuss in later sections, however, homophony alone cannot be the sole determinant of the degree of funniness of a punning joke. Identifying and measuring the contributions of the pun’s semantic features, the joke’s discourse context, and the cultural background and personal humor preferences of the audience would be promising avenues for future research.

2 Background

In incongruity-resolution theories, humor is viewed as two-step process (Uekermann et al. 2007). In the first step, an incongruous element is detected among several incompatible elements, and in the second, this element is linked in a sense-making manner to the context, leading to resolution of the incongruity and a subsequent humorous sensation. Under this interpretation, incongruity is a necessary feature and its resolution a sufficient feature of humor. According to Hempelmann (2003), the questions of whether an element is incongruous in a certain context and whether there is script opposition present is a matter of situation, cultural context, and individual knowledge resources. All of these can vary greatly and influence incongruity to different extents.

Like many others, Attardo et al. (1994) acknowledge phonological similarity to be crucial for incongruity resolution in puns, and use phonological distance as a core measure for the sound-related features in punning jokes. Recognizing that the pun and target word in a punning joke can be similar in sound to varying degrees, Attardo et al. (1994) posit the existence of a threshold in phonological distance that, when reached, makes it impossible to understand a joke. This idea goes at least as far back as Hausmann (1974), who investigated French punning jokes based on the number of differing phonemes between pun and target word and found the largest possible distance to be four phonemes.

Lagerquist (1980) investigated the types of phonemic changes that occur between pun and target words in English puns, among them insertions (additions of phonemes), deletions (omissions of phonemes), mutations (alterations of phonemes), and transpositions (changes in the position of phonemes). Most of the changes were found to involve consonants, and most mutations were focused on a single feature. Since only a quarter of her data showed alterations in the initial segment of the word, she hypothesized a special role for the preservation of the beginning of the sound sequence in the perception of homophony. Other significant observations were the tendencies to preserve the number and stress pattern of syllables between the pun and target. These findings indicate that in the production of punning jokes, the speaker aims to preserve homophony. Keeping in mind the goal of facilitating target recovery, it seems reasonable that the applied changes in the phoneme structure are not major.

In his book on the phonology of puns, Sobkowiak (1991) quantified pun–target sound similarity based on distinctive features and was able to confirm the hypothesis made by Lagerquist (1980) that consonants are more likely than vowels to undergo changes or deletions. He further noted that the understandability of a punning joke increases when the consonant structure is kept intact, since vowels are more mutable and carry less information.

Also in line with previous research, Fleischhacker (2005) states that the pun word needs to be sufficiently similar to the target word for the latter to be recovered. Further, the degree of common representation within a corpus (determined by calculating the ratio of the frequency of a pun–target combination observed in the corpus to the frequency that would be expected if pun types occurred at random, based on the number of relevant word pairs in English) is positively correlated with whether a punning joke is recognized as such in a conversational setting and therefore considered successful. However, according to Hempelmann (2003), while pun–target similarity can explain a higher rate of target recovery, it cannot explain greater funniness. Fleischhacker (2005), on the other hand, assumes that “good” punning jokes with a subtle but quickly recognizable phonological relationship between pun and target are also those that are perceived as funnier.

Jaech et al. (2016) present a study similar to the one described in the present paper. They design a computational model for the recovery of the target word in a punning joke based on the position of the pun word within a sentence context. Rather than calculating phonological similarity between pun and target, they model the transformation probability of the two by averaging the log-likelihood value from their model over the phonemes in the pun, resulting in a phonetic-edit score for each pun–target pair. They use this model to determine whether there is a relationship between the funniness of a punning joke, which they assess through participant ratings, and the phonological similarity of its pun–target pair. Although their statistical analysis finds evidence for such a relationship, their dataset contains only seventeen punning jokes, which calls into question how well their results generalize. A further point of concern is that their funniness ratings were applied with a Likert scale, a practice that has been criticised in the context of humor studies by (inter alia) Shahaf et al. (2015) and Simpson et al. (2019).

Building in part on the notions of joke competence and humor competence (Carrell 1997), Hay (2001) defines a progression of processes associated with the experience of humor – to wit, recognizing, understanding, appreciating, and agreeing with a humorous statement. These processes build upon each other, so that recognition and understanding are necessary for appreciation. Target recovery in punning jokes connects to recognition and understanding in this framework, but it does not necessarily entail appreciation. Our study, which focusses on the perceived funniness of jokes, is connected to the appreciation process, although as Hay (2001) indicates, appreciation is not always possible to disentangle from agreement.

3 Methods

Our study builds on a previously published dataset of punning jokes annotated with information concerning the pun and target and with pairwise funniness judgements. We added further annotations in the form of phonemic transcriptions of puns and targets and used these to calculate four different phonological distance measures between the puns and targets. We then performed correlation analyses between the various phonological distance measures and the funniness ratings. The following two subsections provide further details on the data and distance measures.

4 Results

Figure 1 shows the frequency distribution of normalized, quantized funniness ratings calculated using GPPL across the entire dataset. Most texts scored around 0.6, which is slightly above the median. This normal distribution indicates that most punning jokes were rated in a medium range for funniness with only few of them rated as very low or very high in funniness.

Figure 1:

Frequency distribution of funniness ratings (normalized and quantized).

Our correlation analysis begins with a group comparison between the homographic and heterographic puns. We performed an independent samples t-test comparing mean funniness ratings of the two subsets and found a significant difference in the scores for homographic ( x ‾  = 0.60, σ = 0.14) and heterographic ( x ‾  = 0.52, σ = 0.16) puns (t(2,770) = 13.45, p < 0.0001). The box plots in Figure 2 show a direct comparison between the funniness ratings of the two pun types, indicating that funniness ratings for the homographic puns are higher than those of the heterographic ones. What is striking here is the large variance in funniness ratings in both directions, indicated by the relatively large whiskers for both classes. Further, there seems to be a number of outliers in the lower range of funniness ratings, especially for the homographic group.

Figure 2:

Funniness ratings for heterographic versus homographic puns.

Next, we compare phonological distance measures for the four real-valued distance metrics. Figure 3 shows the frequency distribution of the different phonological distance measures after normalizing and quantizing them. As can be seen in the plot, the majority of puns under investigation are characterized by a low phonological distance between pun and target word. We computed Spearman’s rank correlation to assess the relationship between all four phonological distance measures. There was a positive correlation between ALINE and Levenshtein distance (ρ = 0.91, p = 0.0), ALINE and Covington distance (ρ = 0.83, p < 0.0001), ALINE and Phonetic edit distance (ρ = 0.83, p < 0.0001), Levenshtein and Covington distance (ρ = 0.94, p = 0.0), Levenshtein and Phonetic edit distance (ρ = 0.94, p < 0.0001), and Covington and Phonetic edit distance (ρ = 1.0, p < 0.0001).

Figure 3:

Comparison of phonological distance measures (normalized and quantized).

We now present separate statistical analyses for each distance measure in order to identify possible correlations with the funniness ratings. For this, we once again use Spearman’s rank correlation, finding a significant negative correlation for Levenshtein distance (ρ = −0.089, p = 0.021) as well as for ALINE distance (ρ = −0.165, p = 0.0003). This indicates that lower phonological distance values are associated with higher funniness ratings. For the other phonological distance measures, we observed no significant correlation with the funniness ratings.

Figures 4 and 5 show the respective scatter plots for the negative correlations (indicated by the regression line) between ALINE distance and GPPL funniness ratings, as well as between Levenshtein distance and GPPL funniness ratings.

Figure 4:

Negative correlation between ALINE distance and funniness ratings.

Figure 5:

Negative correlation between Levenshtein distance and funniness ratings.

5 Discussion

Our empirical findings are in line with Fleischhacker’s (2005) oft-repeated assumption that punning jokes in which the pun and target are phonologically closer are perceived to be funnier. In the broader framework of linguistic humor theory, this can be at least partially explained through the concept of target recovery. In particular, theories on the linguistic mechanisms involved in pun processing hold that the implied target word must be (effortfully) recovered by the listener in order for them to understand the punning joke and perceive it as humorous. When the pun and target are closer in sound, it can be easier for the listener to activate the second meaning because there is no need for phonological “stretching.” What remains for the listener is a cognitive exercise on the semantic level in order to evoke the second meaning. Of course, this also holds true for the homographic versus heterographic case, so that target recovery is easier in homographic rather than heterographic puns – not necessarily because of their phonological features but simply also because of the identical orthography of pun and target word.

When comparing mean funniness ratings for homographic and heterographic puns, we found a large variance in both pun types. This raises the question of how meaningful this binary feature is. The fact that raters perceived the funniness of the individual punning jokes very differently might indicate that how funny a larger audience finds a joke cannot be easily measured on a numerical scale – not even when using more reliable and robust approaches that derive numerical ratings from pairwise comparisons.

Our assessment of correlations between the individual phonological distance measures and perceived funniness revealed significant negative correlations for the ALINE and Levenshtein distance measures. Of all metrics we applied, Levenshtein is the most basic; it focusses exclusively on the relative positions of characters without accounting for the phonological information they encode. Its negative correlation with funniness ratings suggests that target recovery may not always involve particularly sophisticated processing. Alternatively, this correlation could be related to the fact that the punning jokes were presented in writing, leading the raters to focus on orthographic rather than phonological or acoustic differences. In contrast to Levenshtein, ALINE distance accounts for specific phonetic features in a fairly sophisticated way. The significant negative correlation with funniness ratings that we observed points to a role in funniness perception beyond mere orthographic similarity, even when the stimuli are presented in writing.

A correlation analysis among our four distance metrics indicated a positive correlation between all of them. This raises the question why only ALINE and Levenshtein distance, but not Covington and the Phonetic edit distance, are significantly correlated with funniness ratings. We suppose that, in the case of Covington, the coarse-grained differentiation of characters into consonants, vowels, and glides is simply not meaningful for accurately judging pun–target similarity, or at least that it obscures the distinctions necessary to make such judgements. With respect to Abydos’s Phonetic edit distance, this measure uses feature weights that are distinct from those derived and explained by Kondrak (2000) for ALINE, and whose provenance is not recorded in the library’s official documentation. Further details on the origin and intended purposes of these feature weights could help explain the null results we observed.

6 Conclusions

This study aimed to investigate the relationship between the phonological distance between pun and target word on the one hand, and funniness ratings on the other hand. Statistical analyses on a large dataset of punning jokes rated for funniness revealed a negative correlation between phonological distance and funniness, providing empirical evidence that punning jokes where the pun and target are closer in sound are generally perceived as funnier.

In a broader humor-theoretic view, punning is based on linguistic ambiguity and is characterized by a resolution of incongruity during the processing of a joke. Successful recovery of the target word is a prerequisite for this incongruity resolution. Lower phonological distance between pun and target is hypothesized to facilitate target recovery, which may explain the higher funniness ratings we observed for more homophonous puns. While the limitations of our data and experimental setup do not allow for firm conclusions concerning the direct causal influence of phonological distance on the perception of funniness, our results are in line with previous findings and lay the groundwork for more in-depth investigations that account for additional factors.