HUMOUR CLASSIFICATION: LAUGHABLE OR SERIOUS BUSINESS?

(1)

HUMOUR CLASSIFICATION: LAUGHABLE OR

SERIOUS BUSINESS?

An interdisciplinary literature overview of linguistic and computational humour classification

Bèta-Gamma - Thema III 26-01-2020

Korijn Moor (11902701) Cilia van Son (11874775) Ghislaine Umuhire (10728031) Marike Knegtering (11700696)

Abstract

Even serious phenomena in life are not untouched by humour. Humour as a complex aspect of life has no consensus definition, and the understanding of humour therefore differs between scientific

disciplines. A common used method to improve our understanding of humour, is the construction of a classification. Humour classifications are among others performed in linguistics and artificial

intelligence. This interdisciplinary research compares humour classifications originating from these disciplines with the use of the Theory of Paradigm by Kuhn. It appeared that classifications

constructed in artificial intelligence are often based on smaller elements like words or phrases, whereas criteria that linguists use to distinguish humour categories take into account emergent phenomena, for example context or structure. With this knowledge, inferring about relationships between classifications can be eased, which could reduce fragmentation between disciplines and enhance a completer understanding of humour. Our advice for future research is to explicitly mention and describe the method that is used to construct a humour classification, since this will ease

(2)

1. Introduction

There is no aspect of life that is not open to humour. Our institutions, heroines, politics, personal affairs, heroes, and everything else can be treated with the same spirit of

playfulness, ridicule, sarcasm and irony. However, we also crave humour: comedians and jokesters are often lavishly rewarded and we seek partners that have a ‘sense of humour’ (Berger,1987; Thorson et al., 1993; Kaslow,1996).

Humour is, however, not only pervasive. The appreciation thereof also differs individually and culturally. Taylor and Raskin (2012) for example claim that humour should be viewed as an universal human faculty that manifests itself variously in different cultures but shares and preserves its basic traits through languages. The question, ‘what is

humour?’, has therefore the ability to trigger debate and controversy.

The difficulty to grasp the concept of humour also translates itself in scientific

research. Knox (1951) writes that “a definition for humour can only be regarded as a lantern

in our hands rather than a formula in our heads” (p. 541). With this he conveys that we

should not hold on too strongly to a definition of humour. Researchers therefore often study humour by looking at its component. They use classifications in order to do this.

Classification is a method by which a phenomenon is arranged in taxonomic groups according to their observed similarities (Costa, Silva & Ribeiro, 2011).

However, research on humour is very fragmented across disciplines. This makes that classifications methods and perhaps classifications can differ per discipline. This research explores how different classification methods relate to one another. More precisely, we will look at humour classification used by linguistics and artificial intelligence as they primarily focus on humour in written form. By doing so, we make our research more feasible as humour and humour expressions can range from gestures and signs to whole poems and stories.

To be clear, this research doesn’t lead to a better understanding of humour itself. It can rather be viewed as a philosophical exercise to understand how humour research in different disciplines relate to one another. This will be done by using the connect method, which means that different meanings of the same concept are being connected. In doing so, important paradigmatic differences can be highlighted, which could be an important stepping stone for further research on humour: humour is very complex and it would be therefore useful if knowledge and insights from different disciplines can be compared and/or integrated.

The structure of this paper will be as follows: we will first provide a literature overview on humour theory and research (Section 1); then we will list and summarize different

classifications of humour used in linguistics and artificial intelligence (Section 2) and finally we compare classifications and classification methods to gain insight in disciplinary

differences (Section 3). In Section 4, we will show the results, and with these results answer in the conclusion (Section 5) our research question: Is there a difference in paradigms

between humour classifications used in linguistic sciences and artificial intelligence? Finally,

in the discussion (Section 6), we will reflect on our research and give some recommendations for future research.

(4)

2. Literature overview

In this section of the report different theories on humour will be discussed. Special attention will be given to linguistics and artificial intelligence, since our research is primarily focused on these disciplines. Furthermore we give some basic theories on humour from philosophy, since these are widely used in humour research in different disciplines.

2.1 Theories on humour

Humour has been described by influential philosophers like Plato, Hobbes and Kant as a negative trait (Morreall, 2012). Contemporary scholars have been mainly been using three theories - incongruity, superiority and relief theory - to categorize the ideas of these

philosophers about humour. What these theories mean according to Mihalcea (2007) is described below.

2.1.1 Superiority Theory

The superiority theory states that humour is the expression of superiority of one over another. This idea can be found in Plato’s Republica, in which he states that humour and laughter should be tightly controlled (Morreill, 2012). He viewed humour and laughter as a way through which people take delight in the fact that they are more virtuous. This theory on humour prompts further philosophical questions about ethics and humour and fits into a large debate about whether art can be unethical. The superiority theory also forms the basis for research in the field of sociology and anthropology in which humour is looked at as tool to regulate social interaction and groups (Mihalcea, 2007).

2.1.2 Relief Theory

The relief theory encompasses that humour enables us to have ‘prohibited thoughts’ that are otherwise censored by ourselves. By making a joke these thoughts can be expressed, giving a feeling of relief. This account of humour was much more described by Freud and

subsequently in psychobiology and psychology (Morreill, 2012).

2.1.3 Incongruity Theory

Psychology also researches humour on the basis of the incongruity theory. The incongruity theory states that humour occurs when two different interpretations are combined in one statement. The first part of the statement creates an expectation, but the second part (outcome) is so absurd or surprising that it creates humour. The incongruity theory is also applied in linguistics, which will discussed in section 2.2.

Paradigm Theory

Kuhn (1974) formulated the idea of a paradigm, which is a collection of all the theories, assumptions, methods and laws a group of scientists shares to practise their “normal” science. The paradigm dictates which topics are interesting to research, and which

(5)

but when they belong to another paradigm, the assumptions and theories they use will be different, and even the outcomes of the research could be different.

2.2 Linguistic approach to humour research

In linguistics there are three prominent theories on humour that evolved from the incongruity theory: firstly the Semantic-Script based Theory of Humour (SSTH), secondly the General

Theory of Verbal Humour (GTVH) and lastly the Ontological Semantic Theory of Humour

(OSTH). The GTVH and OSTH are in fact applications of the SSTH to theories about humour (Raskin, Hempelmann & Taylor, 2009; Mihalcea, 2007).

The SSTH is comparable to the incongruity theory in the sense that humour consists of a build-up, followed by a punch creating surprise. But this SSTH theory elaborates more on the semantics, which is the linguistics and logic concerned with meaning. In the case of the incongruity theory, a humorous text would contain two different interpretations (scripts): one embedded in the build-up and the second one creating the surprise. According to this theory, the opposition between these interpretations is binary. Sources of contrast can for example be: bad/good, positive/negative, actual/non-actual, normal/abnormal and

possible/impossible (Mihalcea, 2007).

The GTVH is a more extended and general form of the previous theory, because it includes other branches of linguistics. It explains that there are six different levels on which a joke can vary. These levels are: script opposition (used in the SSTH), logical mechanism,

situation, target, narrative strategy and language. Despite the extension of this theory, it was

criticized because it would not be falsifiable, and because of a lack of systematic examples (Mihalcea, 2007).

The OSTH combines Ontological Semantic Technology (OST) with the SSTH to create a more powerful way to analyse humour. An OST lexicon consists of the ontological meaning of words and sentences. The aim of an OST is to distinguish the different ambigue meanings each word or sentence has (Raskin et al., 2009, p. 292). For example, the word mouse could be used as meaning a computer mouse, or as the animal mouse. Each meaning of the word gives the sentence in which the word appears another meaning, and thus the context of the story is different. By taking into account the different meanings a word or word group could have, the OSTH theory is more powerful than the SSTH theory in

analyzing the broader context of a humorous text.

2.3 Artificial intelligence’s approach to humour research

The problem of humour in artificial intelligence is twofold. Firstly, we have a problem of detecting humour, and secondly there is the problem of automatically generating humorous things. These problems are interrelated but not the same.

Detecting humour in text is one of the more difficult tasks in artificial intelligence. This is mostly due to the fact that everyone has their own senses of humour. To correctly classify humour in text then, the artificial intelligence has to be trained on what the reader of the text finds humorous (Zhang et al., 2018). This creates a different problem, that is, for a classifier to work correctly one typically needs a lot of data. With different senses of humour the

(6)

classifier needs to be retrained for different persons. This requires a lot of data on all these different senses, which might not be available. This does not stop researches from trying to come up with ways to detect the most general forms of humour. Taylor (2009) and Mihalcea & Strapparava (2006b) describes various methods to achieve detecting humour. West and Horvitz (2019) focused specifically on the task of finding what part of a satirical headline makes it funny and found some aspects of humourful headlines.

Generating humour suffers from the same problem that people have different senses of humour which also plagues the detection of humour. Yet there have been attempts at generating humourful applications: Mihalcea and Strapparava (2006a) created applications which add humourful one-liners to emails and lectures. Stock and Strapparava (2003) investigated how humour could be applied to advertisements. They worked on the HAHAcronym project, which creates funny alternative interpretations to already existing acronyms. Recently, research has been done in how computers could be made to cooperate with humans to be funny. In the research of Wen et al. (2015), a machine suggested funny pictures for humans to use in an online chatting environment. Participants who received help from the computer rated their chat sessions as more fun than participants who did not receive any help. Most of these applications where focussing on specific tasks, and could not create consistently humorous sentences or pictures without a human helping them.

An important thing to note about these researches in the context of our own paper is the product of such humour detection and creating. Mihalcea & Strapparava (2006b) trained a model to detect humour on data which was pre-labeled by humans. In other words the classification of humour was learned by the computer based on what humans thought is humorous. This also means that the computational model itself is, in a way, the classification of humour. Therefore it is quite difficult to pose a general classification for humour in artificial intelligence since most of these models are black boxes, in the sense that they are just numerical values that need to be used in a specific way.

To summarise, humour poses difficulties to standard artificial intelligence approaches of detecting and generating things, due to different personal senses of humour. Yet there have been at least partially successful tries to develop detection and generation applications of humour. A common problem these applications then run into is the fact that what is learned is not easily explainable. The model that was learned is not easily dissected so, the actual classification itself is hidden behind a numerical model that only computers can understand and use.

3. Humour classification

Scientists from disciplines like philosophy, psychology, sociology and anthropology consider humour in research mostly as one phenomenon, whereas computational linguists need to narrow humour down into subcomponents to make it investigable (Dynel, 2009). Linguistics is a very broad discipline that includes multiple approaches and methods, such as

neurolinguistics, sociolinguistics, phonology, morphology and computational linguistics (Aronoff & Rees-Miller, 2001). Because of the broadness in the field of linguistics, humour taxonomies also vary in approach, method and aim. Some taxonomies are based on computational analysis, whereas others are a theoretical list of humour types.

(7)

3.1 Non-computational methods

An example of a theoretical and non-computational classification is proposed by Dynel (2009), who briefly characterises a list of pragmatic types of verbal humour categories, that cannot be reduced to jokes. The aim of this article is to address the variety of conversational humour and familiarise the reader with different types of humour in literature. According to Dynel (2009), it is not desirable to strive for a clear-cut, all-including humour-taxonomy for two reasons: humour categories overlap and merge, which is dependent on the criteria on which an taxonomy is build (1) and taxonomies are subject to continuous expansion due to finding of new second-order and subtypes of humour (2). We will briefly summarize the taxonomy of Dynel (2009) below.

Joke; is thought to comprise a build-up and a ‘punch’ at the end. The punch is considered to be the part that causes surprise or incongruity, which is considered to be a basic

characteristic of humour. Jokes can be divided in subtypes (including one-liners).

Conversational humour; variety of semantic and pragmatic types of humour

● Lexemes and phrasemes; very short, only one or a few words

● Witticism; clever and humorous textual unit interwoven into a conversational exchange, not necessarily of humorous nature, very context-bound in contrast to jokes ○ Stylistic figures ■ Simile/comparison ■ Metaphor ■ Hyperbole ■ Paradox ■ Irony

○ Pun; humorous verbalisation that has two interpretations couched in purposeful ambiguity of a string of words, manifesting itself in one form but conveying two different meanings

■ Idiom ■ Homonymy ■ Polysemy ■ Homophony ○ Allusion

■ Distortion; extra chunks of various lengths are inserted, in any

position, it not only alludes to the source but also entirely changes the meaning of the original formulation, resulting in a humorous effect

● Deletion ● Substitution ● Addition

■ Quotation; direct citations from any pre-existing texts, predominantly popular culture artefacts, which become conversational units available to recipients with sufﬁcient cultural knowledge. Their humorous force stems primarily from the language user’s acknowledgement of the pre-existing text and the quote’s relevance to the situation.

(8)

○ Register clash; using language from an upper or lower class in a new situation

■ Upgrading ■ Downgrading

● Restort; quick and witty response to a preceding turn with which it forms an adjacency pair ○ Subversive ○ Contestive ○ Interactional pun ○ Pragmatic ambiguity ○ Rhetorical question ● Teasing

● Banting; if both parties are willing to engage in a humorous frame, a one-turn tease can develop into a longer exchange of repartees

● Putdown ○ Ridicule ○ Mocking ○ Sarcasm ● Self-denigrating humour ● Anecdote

Taxonomies need a starting point or criterion on which the categories can be build. This starting point differs between different classification and can be theme, subject, cycle, target, origin, narrative form, or length of the joke. As we have seen in the incongruity theory, contrast (or ambiguity) is a main source of humour. For example, Lew (1996) defined different humour categories based on different types of ambiguity, which will we briefly describe below.

● Lexical jokes; lexical refers to words or vocabulary

○ Polysemous lexical items, homonymes and homophones

● Lexicalization of a larger unit (lexico-syntactic); when ambiguity is not caused by one word, but by a larger unit

○ Decomposition of idioms

● Syntactic jokes; syntax = the arrangement of words and phrases to create well-formed sentences in language

○ Syntactic class jokes; two readings differ in terms of syntactic class (and syntactic class)

○ Syntactic function jokes; two readings differ only in terms of syntactic function ● Phonological jokes; ambiguity based on phonology

● Orthographic jokes

● Deictic reference; deictic = a word or expression whose meaning is dependent on the context in which it is used

○ Deictic versus non-deictic interpretation ● Specific versus non-specific interpretation ● Pragmatic ambiguity

● Type of modality

(9)

Another branch in linguistic research uses the approach to focus on lexical devices in

humorous texts. For example, Bucaria (2004) distinguishes three types of ambiguity (lexical, syntactic and phonological), but has focused on the occurrence of ambiguity-based humour in headlines. These types of ambiguity can be divided again in sub-groups.

3.2 Computational methods

Other prominent humour classifications are based on computational methods, as well as in linguistics as in artificial intelligence. Computational linguistics is defined as: “the study of

computer systems for understanding and generating natural language” (Grishman, 1986).

Designed algorithms can be used to test hypotheses and theories proposed by linguists. The distinction between linguistics and artificial intelligence is therefore not discrete.

Humour classification is subjective, since it is dependent on the the expert’s view. This is why automatic humour recognition is a difficult learning task. Nevertheless, automatic computer classification of verbal humour is a very promising direction of research, since it constitutes a fundament for computational humour recognition and offers objective methods1

that can enhance our understanding of humour. Computational classification methods use machine learning techniques including support vector machine (SVM) classifiers, naïve Bayes and less commonly decision trees (Costa et al., 2011).

A mainly used method is the setting up two databases: one with humorous and on with non-humorous texts (control group), on which automatic classification analysis can be used to extract information about distinguishing features (Mihalcea & Pulman, 2007;

Mihalcea & Strapparava, 2006b). The most-used data type for these classification analyses are one-liners (Costa et al., 2011). Also Reyes, Rosso & Buscaldi (2009) used automatic computer analysis to define features that can distinguish humorous one-liners from non-humorous language, to eventually build up to a humour taxonomy. They also inferred about the relative importance of each feature. They used the following features of previous

research (there are more):

● Stylistic features; focusing on adult slang, with the most important feature of sexual information

● Human centric vocabulary; focusing on personal pronouns. This turned out to be the most important word that refers to human-related scenarios. For example, ‘you’ appeared in 25% of the humorous one-liners.

● Human centeredness; focusing on social relationships

● Polarity; focusing on the positive or negative orientation of the data Alongside these features, they also took into account:

● Wh – phrases; focusing on interrogative pronouns; so appearance of question words ● Nationalities; focusing on adjectives of nations

● Keyness; focusing on the extraction of the most representative subjects of the data, so analysis of frequency of used words

● Discriminative items; focusing on the words that belong to a same cluster ● Ambiguity; focusing on the sense dispersion of the words

After the analysis they divided features in two classes: low level features and high level features. Low level features are items that are used to promote humorous situations and are identified as common humorous topics such as sexuality, nationalities, etc. High level

(10)

features are features that are not clearly related to the humorous topic, but are used for producing humour through linguistic strategies (such as polarity and ambiguity). The developed taxonomy is presented in Figure 1.

Definitions (from the taxonomy):

● Stereotypes = humour about ethnic groups ● Pronominal = self-referential humour ● White humour = positive polarity orientation ● Black humour = negative polarity orientation ● Contextual; based on items that denote

exaggeration, incongruity or absurd ● Intra-textual; based on linguistic ambiguity ● Extra-textual; based on pragmatic and cultural

information

Figure 1: A primitive taxonomy of humour (Reyes et al., 2009)

4. Results

In this section we will start with comparing methodologies that are used to construct humour classifications and aims and results of classifications on humour. It will be followed by linking these results to the paradigm theory of Kuhn to explain the differences.

4.1 Classification methods: data use, aim and understanding

After constructing an overview on classifications of verbal humour, we can touch on some notable differences between classifications that are based on different methods. Every methodology has its own possibilities and limitations and as a result, provides its own understanding of humour, which we will explain below. In particular, the difference between computational and non-computational methodologies is outstanding. Computational methods are used in both linguistics and artificial intelligence, but non-computational methods are only used in linguistic sciences.

4.1.1 Data use

For classifications that are constructed by machine learning (computational methods), feasibility and methodological limitations have an impact on the type of data that is analysed and in that way on the taxonomic results too. Most computational analyses are applied to data sets that consist of only one-liners, because these are easier to analyse than larger text

(11)

bodies (Costa et al., 2011). In addition, these classifications distinguish categories using small elements like single words or small group of words, but do not take into account larger aspects that exist on the level of multiple sentences or are based on context (Reyes et al., 2009). As a consequence, the complexity of humour that is considered in these

classifications is limited, creating a bias towards a narrow range of possible classifications. This can also be seen as a difference in the level on which classifications are made: whereas computational classifications are performed on the level of small elements like words or syllables, non-computational methods incorporate higher-level features of humour in language, such as sentences and situations. Classifications based on non-computational methods in general provide a hypothetical or theoretical list of groups of humour, rather than that they quantitatively measure humour characteristics which are afterwards used to build up an taxonomy. However, there are methods that take into account context. This is described by Raskin et al. (2009), who used a computational linguistic method that uses ontological dictionaries to take context into account in verbal humour classes.

4.1.2 Aim

The difference in method also translates itself in a different aim of a classification: the aim of non-computational based classifications is often to familiarize the reader with the variety of humour and improve recognition of types of humour in contexts, rather than an quantitative analysis on the linguistic and structural appearance of humour in language (Dynel, 2009; Reyes et al., 2009).

4.1.3 Understanding

The third and final point that will be addressed, concerns the difference in understanding that the different classifications of humour will provoke, which is a consequence of the

differences in methodology, data and aims. For example, artificial intelligence uses heuristic methods to automatically list features that make it more likely that a sentence is humorous. So one feature is not in itself the element that causes funniness, but can be better seen as a feature that positively relates to the chance that a text is humorous (Reyes et al., 2009). Therefore, this method provides a different understanding of humour than other methods that classify jokes on other criteria.

We tried comparing and describing notorious features in humour classification. However, what is concluded from the addressed parts above is not so straightforward, since the relation between different classification methods remains unclear. Whereas computational methods use initial criteria on which clear-cut distinctions are based, non-computational methods do not explicitly describe how the categories are distinguished. In addition, Dynel (2009), writes that a clear-cut categorisation is not even desirable. In non-computational research, the methods on which the taxonomy is based are not explicitly introduced before the classification is provided. For example, the criteria on which categories are distinguished, are not explicitly mentioned. The method of computational classifications seems to do better with respect to the method of classification, but this type of research has other pitfalls. Computational humour research is focused on only one phenomenon of humour and very little placed in broad-scale theories on humour or taxonomy (Dynel, 2009). Also the contribution of algorithm output on our understanding not explicitly explained.

(12)

The fact that underlying methods and corresponding assumptions are not made explicit, obstructs establishing how one classification is related to another. This is a crucial step in obtaining a more complete understanding of humour, because different

classifications can complement each other. In the Future Research (6.1) section of the Discussion, our recommendations will be made on the basis of these result, but first we will try to explain the results in the light of Kuhn’s Theory on Paradigm and theories on humour from the disciplines linguistics and artificial intelligence.

4.2 Relation of paradigmatic differences to theory

Literature on theories of humour are present in multiple disciplines. Can we use these

theories to explain the difference or relationships of humour classifications described above? Some humour features that are introduced in theories from both linguistics and artificial intelligence seem to recur in the results of humour classification. Ambiguity as main source of funniness has attained much attention. The incongruity theory and the SSTH are mainly based on ambiguity, and GTVH uses ambiguity as a subcategory of humour. This ambiguity recurs in humour classifications. For example, ambiguity and polarity are two key features resulting from computational analysis. Dynel (2009) also writes about the

importance of a build up and surprising ‘punch’ in joke-types of humour, implying that contrast is one way to bring about the humorous effect. Lew (1997) suggests that all jokes are based on ambiguity, which is interesting, since we have seen that some prominent humour taxonomies (and theories) that we have described before, recognise ambiguity as one subcategory. Different classifications thus use different views on humour as a basis and thus bring about a different understanding of humour.

According to Lew (1996), classifications can be single-parameter or either multi-parameter. A parameter can be seen as one dimension of classification. We could use this to explain the systematic differences found between the different classification methods. Whereas computational methods classify in the dimension of single-word use, Lew (1996) for example uses the dimension of ambiguity to distinguish between classes of humour. Multi-parameter research is less performed since it is more complex. Multi-parameter research on humour classification could therefore be seen as an interdisciplinary integration of multiple one-parameter classifications creating a completer humour understanding, as illustrated in Figure 2. Cooperation and communication between disciplines using single-parameter classification could therefore enhance current developed classifications. Clarification of the used dimension and placing classification into context of other classifications are examples of steps that can be taken.

(13)

Figure 2: An interdisciplinary integration of multiple one-parameter classifications on humour (Menken & Keestra, 2016).

Another explanation of the differences between classifications is related to definitions. According to Lew (1997), the appearance of ambiguity in jokes is dependent on how you define ambiguity. Lew (1996) himself defines ambiguity as “the property of a fragment of text

which allows for two or more significantly different semantic interpretations to be arrived at by a substantial proportion of typical text recipients”, but he also mentions that academics

have a right to adopt their own definitions, as long as it does not departs excessively from the previous well-established usages of the term by specialists and non-specialists (Lew, 1996). In conclusion, it seems that there is no consensus definition on ambiguity between disciplines. Formulating consensus definitions on humour concepts eases the integration of multiple one-parameter classifications, because the relationship between classification can be inferred.

5. Conclusion

From the results section we can conclude that there are indeed differences with respect to classifications of humour between linguistic sciences and artificial intelligence. We can see these differences in the scope of data they analyse, the aim of the analysis and and the understanding the analysis gives us about humour.

First, we have observed a difference in scope of data between computational and non-computational methods. There is a certain upper limit on the complexity of analysable data via computational methods due to the limits of computational power available.

Computational analysis are therefore less complex forms of humour, such as one-liners and short jokes. Non-computational methods are not bounded by this problem and often make analysis on higher-level features, for example whole sentences, situations, and context.

Second, the aim of the research methods in both disciplines is different. Where computational analyses are quantitative and more practical, determining the linguistic and structural appearance of humour in language, the intention of non-computational analyses is to familiarise the reader with different forms of humour, and try to improve humour

recognition based on context.

Third, the different disciplines offer a different understanding about humour. Because of the different methods and data sets that are used, the classifications from computational and non-computational research do not match up. And because the underlying assumptions

(14)

that form the fundament of the classifications are not made explicit, it is very difficult to compare these classifications with each other.

Thus we can conclude that indeed there is a noticable difference between these disciplines with respect to humour classifications, and therefore we have detected a difference in paradigms between the linguistic sciences and artificial intelligence.

6. Discussion

To conclude this research we will reflect on our findings and propose some points of consideration as well as direction for further research. We will reflect on the uses and also the limits of this paper and state some recommendations concerning humour research based on our findings.

We mainly focussed on the problem of understanding the disciplinary differences with regards to humour between linguistics and artificial intelligence via humour classification. In other words, we used the tool of classifying humour to analyse the disciplinary differences present in linguistics and artificial intelligence. This gave us an insight in the fragmentation in the field of humour classification. We have found paradigmatic differences between

linguistics and artificial intelligence concerning computational and non-computational methods of humour classification. However, the scope of this paper does not permit us to infer a total picture about fragmentation and paradigmatic differences in linguistics and artificial intelligence as a whole. It might be not unlikely that there are specific areas of these fields which have useful similarities and give opportunities for more practical interdisciplinary research.

Furthermore, humour is a subject not limited to linguistic or artificial intelligence research. Fields like psychology or sociology might also have certain questions, directions of research, methods, and theories regarding the matter of humour. It could be the case that the paradigmatic differences we found in this paper can be united via theories or methods in these other fields. One must therefore consider that the scope of this paper is limited and more research is needed to shed light on the complete picture of humour and humour classifications.

6.1 Further research

As pointed out above we have only considered the disciplines of linguistics and artificial intelligence. Yet humour is a far broader subject which can be observed via much more disciplines. Humour can also be described from for example psychological, sociological perspectives. It could be possible that a broader approach could reveal common grounds in humour classification and be a starting point of unified theories of humour.

In a more different fashion we have seen how methods of making classifications differ from non-computational to computational methods. The computational methods might be enhanced by also using some linguistic or non-computational methods and theories. It could therefore be beneficial to research if it is possible to how to apply linguistic theories to artificial intelligence. It might not be possible to fully integrate non-computational methods into computational ones, yet this result could also shed more light on the complexity of the definition of humour as a whole.

(15)

aware of these differences. This could also provide more information about the damages of this fragmented research areas. It might also indicate to possible solutions of this

fragmentation.

A more specific research topic we observed was ambiguity. It seems that verbal play (humour) often consists of ambiguity. However, ambiguity itself is not per se humorous. For future research, we would therefore recommend more attention to other features of humour and their relationship with ambiguity, to obtain a better understanding of verbal humour.

6.2 Recommendation to the disciplines

Finally, as we have seen in Kuhn's theory on paradigmas, methods always go together with limitations and assumptions that belong together in a paradigm. This is not inherently undesirable, since it is necessary to narrow down a research object and to gain expertise. For example, we found that the field of artificial intelligence does not appear be focussed on theory building and is not able to analyse complex pieces of humour. Yet, the field of artificial intelligence has strong systematic methods to do analysis, something which the field of linguistics does not have. Linguistics then, is very much about theory building and complex analysis, but without the systematic methods, those more complex classifications are also more vague and harder to integrate with other knowledge.

When your aim is to enhance understanding of a phenomenon that crosses the boundaries of multiple disciplines this difference is something that needs to be considered. However, simply not using all interdisciplinary methods in an effort to cross these boundaries would not be preferable and might not even be possible. To be able to integrate knowledge from across these boundaries then, we need a different approach. Writing more explicitly about used methods, criteria and assumptions that go together with the concerned humour classification might create a clearer picture and could provide insights in possible common grounds between classifications. This then may elude us to how different classifications are related and it could provide us with a better and more complete understanding of humour.

(16)

7. References

Aronoff, M., & Rees-Miller, J. (Eds.) (2001). The handbook of linguistics. Oxford, United Kingdom: Blackwell.

Bucaria, C. (2004). Lexical and syntactic ambiguity as a source of humor: The case of newspaper headlines. Humor, 17(3), 279-310.

Costa, J., Silva, C., Antunes, M., & Ribeiro, B. (2011). The importance of precision in humour classification. International Conference on Intelligent Data Engineering and Automated Learning, 271-278.

Dynel, M. (2009). Beyond a joke: Types of conversational humour. Language and Linguistics

Compass, 3(5), 1284-1299.

Grishman, R. (1986). Computational linguistics: an introduction (1st ed.). Cambridge, United Kingdom: Cambridge University Press.

Kuhn, T. S. (1974). Second thoughts on paradigms. The structure of scientific theories, 2, 459-482. Lew, R. (1996). An ambiguity-based theory of the linguistic verbal joke in English (Doctoral dissertation). Adam Mickiewicz University, Poznán, Poland.

Lew, R. (1997). Towards a taxonomy of linguistic joke. Stud. Angl. Posnanien, 31, 123–152. Menken, S., & Keestra, M. (2016). An introduction to interdisciplinary research (1st ed.). Amsterdam, The Netherlands: Amsterdam University Press B.V.

Mihalcea, R., (2007). The multidisciplinary facets of research on humour. International Workshop on

Fuzzy Logic and Applications, 412-421.

Mihalcea, R., & Strapparava, C. (2006a). Technologies that make you smile: Adding humor to text-based applications. IEEE Intelligent Systems, 21(5), 33-39.

Mihalcea, R., & Strapparava, C. (2006b). Learning to laugh (automatically): Computational models for humor recognition. Computational Intelligence, 22(2), 126-142.

Mihalcea, R., & Pulman, S. (2007). Characterizing humour: An exploration of features in humorous texts. International Conference on Intelligent Text Processing and Computational Linguistics, 337-347.

Perks, L. G. (2012). The ancient roots of humor theory. International Journal of Humor Research,

25(2), 119-132.

Raskin, V., Hempelmann, C. F., & Taylor, J. M. (2009). How to understand and assess a theory: The evolution of the SSTH into the GTVH and now into the OSTH. Journal of Literary Theory, 3(2), 285-311.

Reyes, A., Rosso, P., & Buscaldi, D. (2009). Evaluating Humorous Features: Towards a Humour Taxonomy. IICAI, 9, 1373-1390.

Stock, O., & Strapparava, C. (2003). Getting serious about the development of computational humor.

IJCAI, 3, 59-64.

Taylor, J. M. (2009). Computational detection of humor: A dream or a nightmare? The ontological semantics approach. 2009 IEEE/WIC/ACM International Joint Conference on Web Intelligence and

(17)

Wen, M., Baym, N., Tamuz, O., Teevan, J., Dumais, S. T., & Kalai, A. (2015). OMG UR Funny! Computer-Aided Humor with an Application to Chat. ICCC, 86-93.

West, R., & Horvitz, E. (2019). Reverse-Engineering Satire, or “paper on computational humor accepted despite making serious advances”. Proceedings of the AAAI Conference on Artificial

Intelligence, 33, 7265-7272.

Zhang, D., Song, W., Liu, X., Liu, L., & Zhao, X. (2018). Research on Humor Recognition. 2018

HUMOUR CLASSIFICATION: LAUGHABLE OR SERIOUS BUSINESS?