Theories of Verbal Humor

Abstract

In this paper, we propose a methodology that aims to develop a recommendation system for jokes by analyzing its text. This exploratory study focuses mainly on the General Theory of Verbal Humor and implements the knowledge resources defined by it to annotate the jokes. These annotations contain the characteristics of the jokes and hence are used to determine how alike the jokes are. We use Lin’s similarity metric and Word2vec to calculate the similarity between different jokes. The jokes are then clustered hierarchically based on their similarity values for the recommendation.

Finally, for multiple users, we compare our joke recommendations to those obtained by the Eigenstate algorithm which does not consider the content of the joke in its recommendation.

Keywords: Computational humor; General theory of verbal humor; Clustering; Joke similarity

Introduction

Humor is an interesting phenomenon that can be identified most of the time but is very difficult to ‘define’ (McGhee & Pistolesi, 1979). Yet, its importance becomes more evident with humorless technological advances.

Humor is much more than just a source of entertainment; it is an essential tool that aids communication. Various empirical findings have confirmed that stress and depressing thoughts can be regulated with the help of humor (Francis, Monahan, & Berger, 1999). Positive psychology, a field that examines what people do well, notes that humor can be used to reduce tension, make friends, make others feel good, or to help buffer stress (Lurie & Monahan, 2015) (Ruch & Heintz, 2016).

The need for humor in a computerized setup is often discussed and many researchers have presented their findings.

Get quality help now

Bella Hamilton

Verified

Proficient in: Humor

5 (234)

“ Very organized ,I enjoyed and Loved every bit of our professional interaction ”

+84 relevant experts are online

Hire writer

Some of the applications of computational humor are human-computer interfaces (Morkes, Kernal, & Nass, 1998), education (McKay, 2002), edutainment (Stock, 1996), understanding how human brain works (Binsted et al., 2006; Ritchie, 2001), etc.

The advancements in AI have sowed the seeds of the idea that computers can understand the human language. Since humor is a ubiquitous aspect of the human experience, it is fair to expect the computers to take into consideration the humorous facet. Almost two decades ago, it was pointed out that if computer systems can incorporate humor mechanisms, then these systems would appear to be more user-friendly hence less alien and intimidating (Binsted, 1995). This statement still holds and to achieve this, one of the key things to consider is that different people find different things funny which makes research in this field both challenging and interesting.

Verbally expressed or verbal humor is a common form of humor, and one of the subclasses of verbal humor is the joke. A joke can be defined as “a short humorous piece of literature in which the funniness culminates in the final sentence” (Hetzron, 1991). This paper focuses on verbally expressed humor with the help of jokes.

The motivation for this research comes from the observation that the smart assistants like Alexa and Siri recite the same jokes to all the users without considering their humor preferences. The idea behind this research is to come closer to understand human humor preferences and recommend jokes based on it. We propose a framework to recommend jokes to the users by taking into account the text of the joke as well as the liking of the users. Our assumption is that individuals like certain categories or types of jokes. These types can be identified through the individual’s funniness ratings.

This framework is centered on the identification and quantification of similarity between jokes. The General Theory of Verbal Humor states that jokes can be represented and compared with the help of six knowledge resources (Attardo & Raskin, 1991). We use these knowledge recourses to find joke similarity in the Jester Dataset. Once similar jokes are identified, we explore whether subject ratings confirm the similarity.

There exists a joke recommendation system, Jester, (Goldberg, Roeder, Gupta, & Perkins, 2001) but it considers the users and the text of the joke as a black box and relies solely on the user ratings for the recommendation. It works as a baseline model to our proposed model and we compare the joke recommendations to the same user by both the models. We also analyze the ratings given by the users to the jokes that are considered similar to our model.

Humor Theories

Humor studies date back to the era of Plato (Philebus) and Aristotle (Poetics). There are three major classes of humor theory: superiority theories, release/relief theories, and incongruity theories. The general idea behind superiority theories was that people laugh at other people’s misfortunes since it makes them feel superior to them (Attardo, 1994) (Raskin, 1985). Release/relief theories assert that humor and laughter are a result of the release of nervous energy (Meyer, 2000). The family of incongruity theories states that humor arises when something which was not anticipated happens (Raskin, 1985). There has been a debate among various thinkers if incongruity alone can be considered to be sufficient enough to be able to mark something as funny (Suls, 1977).

This gave birth to the Incongruity-Resolution theories which focused not only incongruity but also on its realization and resolution. Suls (1972) proposed a two-stage model that stated that when there is some incongruity in the text, if one can resolve it then it’s a joke otherwise the text leads to puzzlement and no laughter (Ritchie, 1999). Another model to resolve incongruity was summarized by Ritchie (1999) as the surprise disambiguation model which states that the setup of the joke has two different interpretations out of which one is more obvious than the other. The hidden meaning of the text is triggered once the punchline is reached.

Raskin’s Script-based Semantic Theory of Humor (Raskin, 1985) is the first linguistic theory of humor. It is regarded as neutral concerning the three classes of humor theories. SSTH states that a joke carrying text should be fully or partially compatible with two scripts and these scripts must oppose. Raskin introduced several types of script oppositions, such as real/unreal, actual/non-actual, good/bad, life/death, sex/non-sex. The following joke is analyzed in Raskin (Raskin) with the scripts of Doctor and Lover being the two scripts that overlap and oppose.

Joke1: “‘Is the doctor at home?’ the patient asked in his bronchial whisper. ‘No,’ the doctor’s young and pretty wife whispered in reply. ‘Come right in.’”

The joke evokes the script of a Doctor due to the words “doctor”, “patient” and “bronchial”. The second script, Lover, is triggered by the words “no” as well as the description of the doctor’s wife. The wife’s reply is incongruous to the first script, and thus the second script emerges, which makes the punchline, “come right in” explainable. The joke is said to have a partial script overlap between Doctor and Lover – both scripts contain a person that comes to the doctor’s house for a visit – and since these scripts are opposing each other based on sex/non-sex, the text is considered a joke (Attardo, 1994) (Raskin, 1985).

Attardo and Raskin (1991) revised the SSTH into General Theory of Verbal Humor which stated that the jokes can be described using six knowledge resources (KRs) which are ordered hierarchically: script overlap/opposition (SO), logical mechanism (LM), situation (SI), target (TA), narrative strategy (NS), and language (LA). Upon empirical verification of the KR hierarchy, LM was found to behave differently than predicted (Ruch, Attardo, & Raskin, 1993). GTVH also made the comparison of jokes possible with the KRs. The higher the number of common parameters in jokes, the higher is joke similarity. Additionally, jokes that differ only in SO are less similar than the jokes that differ only in LM, than the jokes that differ only in SI and so on. For example, the following jokes are introduced in Attardo & Raskin (1991) to illustrate the comparison:

Joke2: “How many Irishmen does it take to screw in a light bulb? Five. One to hold the light bulb and four to turn the table he’s standing on.”

Joke3: “How many Poles does it take to wash a car? Two. One to hold the sponge and one to move the car back and forth”.

Joke4: ‘Do you think one Pole can screw in a light bulb?’ ‘No.’ ‘Two?’ ‘No.’ ‘Three?’ ‘No. Five. One to screw in a light bulb and four to turn the table he’s standing on.’

The KRs representing these jokes are represented in Table 1:

Table 1: Joke Comparison (Attardo & Raskin, 1991)

KR Joke3 Joke4 Joke5

SO Dumbness Dumbness Dumbness

LM Figure-Ground Reversal Figure-Ground Reversal Figure-Ground Reversal

SI Light Bulb Car Wash Light Bulb

TA Irish Poles Poles

NS Riddle Riddle Ques -Ans

LA LA 1 LA 1 LA2

Here, jokes 3 and 4 differ in three of the parameters, namely, LA, NS, and SI; jokes 2 and 3 differ in two of them, namely TA and SI; and jokes 2 and 4 in three of them, namely LA, NS and TA. Jokes 2 and 3 are the most similar since they differ in only two knowledge resources. Since SI is placed at a higher level in the hierarchy, jokes 3 and 4 are the least similar even though they have the same number of different KRs as jokes 2 and 4. This paper will rely on this theory to process humor computationally.

Methodology

We assume that previously unseen jokes should be recommended to users as well as jokes that have been rated by others (and thus, have been seen by the system). This means that the content of the jokes, not just the user ratings, has to be taken into consideration. To do so, we develop a methodology to compare jokes based on their content, find their similarity, and then cluster them accordingly. The jokes which are clustered together — and have at least one highly rated joke – serve as the recommendations for the users.

Corpus

This paper adopts jokes from the Jester dataset. We use version 3 of the dataset which is an updated dataset of the previous versions. Version 1 has rating values from -10 to +10 of 1000 jokes collected between April 1999 to May 2003. It also includes 50 new jokes which were not in dataset 1. Apart from this, it has an updated version of the dataset’s version 2 with 115,000 new ratings.