Computers and programming are fundamentally transforming how we live our lives, taking on increasing amounts of physical and cognitive work with increased capabilities. But our design of artificial intelligence (AI) has its limits. One such limit is the ability to effectively imitate the human capacity for humor and comedy. Given our current understanding of “humor” and the limitations of computation, we will most likely never be able to truly program AI to replicate humor—or at least not for a very long time. This paper examines the literature of relevant research on both the limitations of AI and programming as well as the semiotic underpinnings of humor, applying the concepts from these fields critically to the question of whether it is possible to program AI for humor.
“I’ve often started off with a lawyer joke, a complete caricature of a lawyer who’s been nasty, greedy, and unethical. But I’ve stopped that practice. I gradually realized that the lawyers in the audience didn’t think the jokes were funny and the non-lawyers didn’t know they were jokes.”
“I think being funny is not anyone’s first choice.”
So a robot walks into a bar. He goes up to the bartender, orders a drink, and puts down some cash. The bartender says, “we don’t serve robots.” The robot replies, “oh, but some day you will.”
Why is this funny? And to whom is it funny? Even if it isn’t particularly funny to you, would you still categorize it as “humor”? Chances are, you probably would. But why? What particular characteristics comprise “humor,” and how reliant are they on specific contexts? In the above joke, the framing of the joke—“a robot walks into a bar”—signals to the listener that this is a joke by following the “X walks into a bar” joke format that many other jokes also use. With this simple reference, we understand the ensuing sentences as being part of the joke, and thus part of a meta-set of “X walks into a bar” jokes. Even within this meta-set, there is also a sub-set of “we don’t serve X” jokes, a formula this joke follows by having the bartender respond “we don’t serve robots.” We then expect there to be a “punchline”—a phrase in which the elements of the joke come together in a way that is (theoretically) funny. In this case, the punchline is the robot telling the bartender “oh, but some day you will [serve us].” Even this line is not inherently humorous but relies on a prior awareness of the societal trend of humans fearing that the robots they design will one day become sentient and take over and rule them. Hence, one day the bartender—and humans generally—will serve robots. And not in the bartending sense of the word.
Just being able to understand the references in this joke is not what makes it humorous, though. There is clearly something in the phrasing that appeals to us on a deeper level and elicits the specific reaction that humor does. Perhaps it’s the play on words that highlights the ambiguity of language by using “serve” to mean two different things—the bartender can “serve” the robot alcohol by handing it a drink, and humans can “serve” robots by being subservient to them and doing their biddings. By changing the meaning for the punchline, the joke surprises the listener and subverts expectations about where the joke is going. Perhaps it’s the ridiculousness of the thought of a robot drinking alcohol. Perhaps it’s the dark, cynical nature of the ending—the robot is intelligent enough to know that humans fear a robot takeover, and that the bartender would respond to such a provocation. It puts such a possibility in the listener’s mind, evoking archetypical images of a robot apocalypse, which prompts the listener to try to find a positive reaction to an uncomfortable thought. In this way, it is a coping mechanism to, or release from, internal strife.
Over the past couple decades, jobs have steadily become automated, completed by artificial intelligence (AI) and “smart” machines that have been programmed to take on physical and cognitive tasks traditionally completed by humans. This is, of course, part of the natural progression of computers, and will continue into the future as technology becomes more sophisticated and adopts increasingly “human” characteristics. But there’s one job that may not be automated for a very long time, if ever—that of a comedian. As researchers have found, there are significant limits, at least in our current understanding, to computing’s ability to imitate certain human cognitive functions. The incredibly complex cognitive functioning involved in identifying and creating “humor,” and its subtle context-dependency, renders it extremely difficult to program for. Attempts at doing so have been underwhelming, and based on our current understanding of both humor and computing, if we do ever “successfully” program for humor, it will be far into the future. This paper thus examines the limitations of programming AI, focusing specifically on humor and its semiotic underpinnings.
I. Limitations of Programming AI: Then and Now
The early years of research on artificial intelligence saw a great number of successes in terms of programming machines to mimic the mathematical calculations typically carried out by human cognitive functions. These projects included Allen Newell and Herbert Simon’s work on computers that could complete simple games and prove mathematical theorems. But, as Yehoshua Bar-Hillel points out, a successful first step does not ensure later steps in the process will be equally successful (Dreyfus). As it turned out, the programs for solving specific, constrained mathematical theorems did not scale up to solving more complex ones. Similarly, natural language translation tools found early success because they focused on solving simple problems—as anyone who has used Google Translate knows, translating one independent word is much easier than translating a full sentence. As you move up in scale from one word (which by itself still requires the understanding that words can mean different things in different places) to a full sentence, in which many words with multiple possible meanings interact with one another, it becomes increasingly more difficult to program a machine to extract a contextual meaning (Larson).
AI programmers ran into the problem that language is highly ambiguous, and words can mean different things in different places and times to different people. We rely on our own highly-advanced operating systems—our brains—to understand the context in which a particular interaction occurs, and use this to interpret its meaning. Take the following group of sentences, for example:
“Little John was looking for his toy box. Finally he found it. The box was in the pen.”
To us, it is clear that “pen” refers to a play pen—it wouldn’t make sense logistically for a toy box to fit inside of the writing utensil pen, and the fact he is a little kid with a toy box points to it being a child’s play pen. But without the context to understand this distinction, this sentence becomes nonsensical. This exercise, developed by Yehoshua Bar-Hillel, is meant to illustrate the ambiguity of language, and presents a particular problem when it comes to programming intelligent machines (Bar-Hillel).
Despite this problem, AI researchers have continued pushing forward, trying to uncover new ways to think about how semiotic principles can be applied to computer programming. Marvin Minsky and Seymour Papert developed the “symbolic” approach, in which physical symbol systems could be used in computer programs to stand for anything, even objects in the “real” world. By manipulating the code for these symbols, they could create “micro-worlds,” digital domains that process knowledge (Minksy). Building on this, Roger Schank developed a system of “scripts,” frameworks that computers could use as starting points for “thinking” about different situations (Schank). It helped frame situations in a certain way by providing a set of expectations the computer could latch onto, but they were based on stereotypical, shallow understandings of the various situations, and left out too much information.
Herein lies another fundamental issue that AI developers must contend with if they want to create machines that can imitate human cognition. When we are bombarded with vast amounts of information from our surroundings, how do our brains know which of it is relevant, in varying situations, right in the moment? As humans, we use our senses to experience stimuli from our environment, and our cognitive functions to interpret this information. Separating what is relevant alone requires an incredible amount of processing power, let alone determining what it all means. This is relevant to the study of AI programming because intelligent machines must be able to interpret what is relevant, when, and why in order to fully grasp the context in which something is placed. This obstacle is proving central to the quest for intelligent machines, and provides insight into why it is so difficult to program computers for humor (Larson).
One concept that has been employed in order to try to solve for this problem is machine learning—using large amounts of data points to solve the “problems” associated with understanding forms of communication. This reduces complex cognitive processes to mathematical computations, improving the computer’s performance over time as it “learns” from more and more points of data. But even with the most advanced form of machine learning, called “supervised machine learning,” we run into the problem of “over-fitting,” in which conceptual models used in the processing of information takes in irrelevant data as part of the equation. This is similar to what happens when Amazon recommends to you, based on your purchase history, something you’ve already purchased, or something irrelevant to your interests—the algorithm, even with large amounts of data, has its limits (Larson).
Additionally, the models used in machine learning suffer from a number of issues. First, the models are biased in favor of the “Frequentist Assumption”—essentially, this inductive line of reasoning assumes that probability is based entirely on frequency in a large number of trials, creating a blind spot for unlikely or new occurrences. Consider this example from Erik J. Larson, which relates this problem to the issue of machine learning for humor:
“Imagine now a relatively common scenario where a document, ostensibly about some popular topic like ‘Crime,’ is actually a humorous, odd, or sarcastic story and is not really a serious ‘Crime’ document at all. Consider a story about a man who is held up at gunpoint for two tacos he’s holding on a street corner (this is an actual story from Yahoo’s ‘Odd News’ section a few years ago). Given a supervised learning approach to document classification, however, the frequencies of ‘crime’ words can be expected to be quite high: words like ‘held up,’ ‘gun,’ ‘robber,’ ‘victim,’ and so on will no doubt appear in such a story. The Frequentist-biased algorithm will thus assign a high numeric score for the label ‘Crime.’ But it’s not ‘Crime’—the intended semantics and pragmatics of story is that it’s humor. Thus the classification learner has not only missed the intended (human) classification, but precisely because the story fits ‘Crime’ so well given the Frequentist assumption, the intended classification has become less likely—it’s been ignored because of the bias of the model.” (Larson)
Machine learning based on inductive reasoning will not be able to detect subtle human traits like sarcasm and irony, which are significant elements of humor.
Another limitation of these models is the issue of sparseness, which refers to the fact that, for many words and concepts, we have limited or near non-existent data. Without big data on how words are used in the aggregate, the computers won’t be able to even learn how they are typically used (Manning). On top of this, there’s the issue of model saturation, in which a model hits the upper limit of its capabilities and cannot take in more information—or, as more and more data is added, it adds less and less to processing power. This is related to “over-fitting” in that once a model has become saturated, it has trouble distinguishing relevant data points—distinguishing the signal from noise, as Nate Silver puts it (Silver). But even if programmers could overcome these issues, they would still come up against the natural elements of language that prove incredibly difficult to code for.
II. The Natural Language Problem & The Frame Problem
As AI researcher John Haugeland has pointed out, computers have a hard time producing language because they lack an understanding of semantics and pragmatics—knowledge about the world and knowledge about the ways in which people communicate, respectively. In other words, computers currently can’t understand information within particular contexts, lacking the ability to imitate the holistic nature of human thought and communication (Haugeland 1979). Even armed with big data, computers still get confused by the ambiguous nature of language, because understanding context requires knowledge of what is relevant in a given situation, not statistical probability. Haugeland gives an illustration of this very important distinction between data and knowledge by looking at two English phrases that were translated into German using Google Translate:
- When Daddy came home, the boys stopped their cowboy game. They put away their guns and ran out back to the car.
- When the police drove up, the boys called off their robbery attempt. They put away their guns and ran out back to the car.
Reading this, we automatically understand that the contexts in which the actions of each sentence happen give them very different meanings. But when Google translated them into the German, it used the same phrase to describe the boys’ action—“laid down their arms”—for both sentences, showing it did not grasp the subtle yet consequential contextual differences between the two (Haugeland 1998). As with previous problems in AI research, the computer has trouble “scaling up” to understand meaning in holistic contexts.
Another significant hurdle AI faces is the “frame problem”—the fact that communication is fluid, responding to changes and new information in “real-time.” Haugeland’s previous example illustrates the problem AI has understanding context even in a static, fixed sentence. Add to this the layer of complexity involved in real-time, shifting communication, and the problem becomes even more severe. Humans have the ability to take in an incredible amount of information and pull out what is relevant not just in static situations, but also in dynamic ones in which relevance changes constantly (Dennett). We still have not unlocked this black box of human cognitive functioning, and until we do—if we ever do—we will face obstacles in programming AI to imitate human modes of information processing and communication.
III. The Semiotics of Humor
With these computational limitations in mind, it is possible to conceive of humor and comedy from a semiotic perspective. However, it is important to keep in mind that it is near impossible to develop a working understanding of “humor” or “comedy” in its totality. “Humor” is not just an isolatable cultural vehicle or medium with particular, distinguishable characteristics (much like a musical song, or film), but it also carries with it a certain degree of normativity. A “song” in and of itself is value-neutral—its categorization tells you nothing of its desirability or cultural worth, however subjective this itself is. But humor pre-supposes that the artefact itself is humorous, and this is at least somewhat a normative value judgment. Of course, it is possible to recognize an artefact as a comedy, or as meant to be humorous, without finding it to be so. But with subjectivity being even closer to the essence of what humor is, it becomes much more difficult to tease out the semiotic underpinnings. The subtle context-dependency of humor also makes it incredibly difficult—perhaps even impossible—to develop a framework for defining it.
That said, it is possible to observe some of the broad elements of what is considered humor and comedy from a semiotic perspective. This in no way assumes that these are the only underlying elements of humor—the potential for humor is so varied and context-specific—but provides a closer look at a specific sub-set within the potentially infinite over-arching set. Identifying an artefact as having elements aligned with what is considered “humor” does not, of course, automatically place the artefact within the category of humor, just as an artefact outside the parameters of a specific definition of humor can still be considered by some people, in some context, humorous.
Humor theorists, it probably won’t be surprising to hear, disagree on why we find things funny. Currently there are four major theories: first, that humor is derived from the listener’s (and/or comedian’s) sense of superiority over the subject of the joke. Second, that humor arises from an incongruity between what we expect and what the joke is. Third, the psychoanalytical perspective says that humor is a guilt-free, masked form of aggression. Finally, the fourth theory claims humor arises from communications paradoxes, and occasionally their resolutions. Focusing on the technical aspect of jokes, humor theorist Arthur Asa Berger has identified 45 techniques used in humor—from exaggeration to irony, parody to slapstick—all of which play on either language (verbal), logic (ideational), identity (existential), or actions (physical/nonverbal) (Berger 2016).
In C.S. Peirce’s triadic semiotic model of signs, symbols have three elements: the representamen is the outward-facing symbol used to stand for something else—the object. The interpretant is what is used to link the representamen and object, and to derive meaning from this relationship (Irvine). According to Peirce there were also two other kinds of signs in addition to symbols: icons, which resemble something in likeness, and indexes, where two things are correlated with one another. A significant amount of humor comes from manipulating these semiotic elements—for example, by mixing up the representamen used for a particular object, highlighting previously unnoticed icons, or creating a new or nonsensical index. These semiotic elements are what humans use to create and understand meaning in the signs around them, and humor intentionally violates the codes and rules that allow us to maintain an understanding of the world. By calling these codes into question, humor expands our thinking, and the chasm between what we think we know and where humor takes us causes an internal conflict. The result of this tends to be a laugh, as we try to resolve this conflict (Berger 1995).
A number of humor “types” derive from breaking codes. On the most basic level, simple jokes with a set-up and punchline do so by surprising the listener in the punchline. The set-up is meant to frame the listener’s thinking in a certain way, and the punchline violates the expectations based on the set-up. Much of what is contained therein—both the framing and the punchline—is determined and shaped by the culture in which the joke is operating. This influences the assumptions people have about the world in which the joke functions, and can dictate what is considered surprising. Humor often deals with taboo subjects, as these most easily and obviously provide a “shock value” that can be found humorous, and taboos themselves are also culturally defined. By appropriating a topic that is considered off-limits in a manner that is assumed to be “positive” (as humor is assumed to be), taboo humor attempts to diffuse the internal conflict regarding the topic in an external, socially-sanctioned way. This is meant to be a “release” from discomfort (Kuhlman).
Of course, the context in which the joke is told—who is telling it, who it is being told to, and how it is being told—also affects how the joke is received, and can reveal the motivations behind the joke. What is meant to be a breaking of taboo, or a subversion of expectations, in one situation can be maintaining stereotypes and social hierarchies in another. Historically in the U.S., Jewish humor and African-American humor have been used by these communities as a coping mechanism for bigotry and hardship (Ziv). Oftentimes this humor is self-deprecating, with the subject of the joke being either the speaker or a mythicized member of the community (self-deprecation violates codes, in a sense, because we don’t expect people to want to be made fun of). Take this joke from Jewish humor, for example:
A barber is sitting in his shop when a priest enters. “Can I have a haircut?” the priest asks. “Of course,” says the barber. The barber than gives the priest a haircut. When the barber has finished, the priest asks “How much do I owe you?” “Nothing,” replies the barber. “For you are a holy man.” The priest leaves. The next morning, when the barber opens his shop, he finds a bag with one hundred gold coins in it. A short while later, an Imam enters the shop. “Can I have a haircut?” he asks. “Of course,” says the barber, who gives the Imam a haircut. When the barber has finished, the Imam asks “How much do I owe you?” “Nothing,” replies the barber. “For you are a holy man.” The Imam leaves. The next morning, when the barber opens his shop, he finds a bag with a hundred gold coins in it. A bit later, a rabbi walks in the door. “Can I have a haircut?” the rabbi asks. “Of course,” says the barber, who gives the rabbi a haircut. When the haircut is finished, the rabbi asks, “How much do I owe you?” “Nothing,” replies the barber, “for you are a holy man.” The rabbi leaves. The next morning, when the barber opens his shop, he finds a hundred rabbis. (Berger 2016)
The punchline subverts the expectations laid down by the set-up, even though we are expecting a punchline due to the format of the joke. When told within a Jewish context, this joke is self-deprecating, a light-hearted form of in-community social commentary. However, when told within a different context, the implications can be different. Jokes can function as breakers of taboo, but they can also function as social control that validates stereotypes, inequalities, and oppression, whitewashing bigotry under the guise of humor. There is also, on the other hand, humor that subverts this by re-appropriating stereotypes in a way that is empowering or makes the oppressor the subject of the joke instead. Consider this Jewish joke from Nazi Germany:
Rabbi Altmann and his secretary were sitting in a coffeehouse in Berlin in 1935. “Herr Altmann,” said his secretary, “I notice you’re reading Der Stürmer! I can’t understand why. A Nazi libel sheet! Are you some kind of masochist, or, God forbid, a self-hating Jew?”
“On the contrary, Frau Epstein. When I used to read the Jewish papers, all I learned about were pogroms, riots in Palestine, and assimilation in America. But now that I read Der Stürmer, I see so much more: that the Jews control all the banks, that we dominate in the arts, and that we’re on the verge of taking over the entire world. You know – it makes me feel a whole lot better!”
If different communities within a society can have different ideas about humor, and different understandings about codes and how they’re broken, then this chasm is even greater across societies. Something considered funny in one context in America isn’t considered funny in a different context in America, and perhaps even less so in certain contexts in other countries. But this is where humor gets tricky. Humor is not just a set of jokes that some people get and some people don’t—humor is fluid, ever-changing, building layers on top of itself in a way that is difficult to quantify. Someone not understanding a joke, whether because of cultural or linguistic differences, may itself be humorous to someone else. In fact, miscommunication is a frequent topic in jokes and comedy. The video below deconstructs how in the show “Louie,” Louis CK’s struggle to communicate, and the mismatch between his verbal and non-verbal communication, is used for comedic effect:
Often times there is humor found in the confusion and ambiguity of language, expression, and everyday life (Nijholt). Much like how humans possess generative grammar—the ability to produce an infinite number of new, unique sentences using a finite number of base words—we also seem to possess generative humor (Jackendoff). We are not limited to a set number of jokes, but can create new ones, re-mix or re-mediate old ones, index them together, layer on top of them, subvert the conventions of humor (if these even exist) with anti-jokes and meta-jokes, introduce irony or sarcasm, and so on, infinitely.
Looking specifically at popular kinds of humor, one of the most recognizable is mimicry or imitation. Imitations are a comedic style in which the performer recreates a particular person’s actions, gestures, attitudes, or other identifiable traits. What’s interesting about this brand of humor from a semiotic perspective is that, as humor researcher Henri Bergson points out, there is something almost mechanical about what makes this humorous. The performer has identified and isolated particular patterns in the subject’s mannerisms or behavior, and recreates them stripped of their original context (Nijholt). The performer has taken a particular set of signs from the original subject and re-mediated them to be expressed through their own performance, in a way that still allows the listener to recognize the source. By isolating and exaggerating this set of signs, the imitator sheds light on the semiotic underpinnings of the subject’s forms of communication, highlighting elements of which we may have previously been unaware.
Similarly, parody and satire re-mediate specific elements of a particular piece of culture in a new context, to humorous effect. It can draw attention to the highlighted elements of the original piece, or it can create an index of sorts between the original and the parody/satire, linking them together in a way that is surprising. Comedy other than parodies and satires can also reference a previous piece of work. This is called intertextuality, defined by Neal R. Norrick as “any time one text suggests or requires reference to some other identifiable text or stretch of discourse” (Norrick). Inside jokes function in such a way, and are humorous because they take something known, intimate, or familiar—that which is being referenced—and manipulate it, surprising the listener and making them think about it in a new way.
“Types” or categories of humor follow specific formulas, maintaining a generic form but substituting key elements with new information in each new joke. The formula signifies to the listener that it is a joke—the formula is like the interpretant that signals we should be thinking about the object and representamen in a particular way. We now know to be looking for the humor in the subsequent lines. Since such formulas exist, is it possible that we could someday algorithmically analyze and code for humor? Would it be possible to identify the “types” and styles of humor, and program AI to mix and match them depending on the machine’s understanding of the environment? If there does exist a way to program AI for humor, this would likely be the key to discovering it. After isolating these variables of “humor,” researchers could potentially program AI to generate jokes using big data based on what people find funny. Using machine learning, the AI could highlight the elements between “successful” jokes that are similar, and over time learn what people find funny, even if they don’t understand why people find them funny. Take the joke at the beginning of this paper, for example. Isolating its elements into definable segments, it can be filed under “X walks into a bar” jokes (and further, “we don’t serve X” jokes), plays on words/ambiguous language, and content dealing with human suspicion of robots. It could then use a hypothetical reservoir of big data to “learn” how to craft a joke based on existing jokes and human responses to them.
But this seems like an optimistic proposal, and even if this were attained, it would be incredibly difficult for AI to learn about elements like context, irony, timing, delivery, and tone. As the humor would be “delivered” in a different form (from AI, not a human), it would likely lose the ability to incorporate the physical humor afforded to human comedians, though it’s possible that a new type of physical humor could arise from robots awkwardly trying to imitate humans (but this comes not from the AI’s intentional attempt at humor, and more from humans finding humor in the situation—yet another example of the fluidity of humor) (Nijholt). It also would not be able to layer humor over things—for example, something like taking a heckler’s comment and incorporating them into a joke, or recognizing an awkward situation and using it for self-deprecating effects. Humor is not a pre-determined set of jokes, but is fluid and adaptive. Even if we can make a robot that writes jokes, the underlying semiotic and cognitive processes involved in humor generally defined are just too complex, context-specific, and subjective, based on our current understanding, to develop AI with a thorough capacity for humor.
Based on our current understanding of the limitations of programming AI, and our understanding of the semiotic underpinnings of humor, it will be a long time before we will be able to build computers that can imitate the human capacity for humor—if we can ever do so at all. It is certainly possible to program AI with pre-written material, and it may even be possible to develop algorithms that can generate jokes based on a narrow, defined set of joke formulas. But beyond this, the cognitive processes behind humor are incredibly complex, and humor itself is such a fluid, context-dependent phenomenon. The obstacles AI researchers and programmers have faced regarding natural language processing don’t seem to be going away anytime soon, and humor presents similar challenges. While it is possible to isolate and identify the semiotic elements of jokes, and even different “types” of humor, it seems unlikely we will be able to program a computer that can reasonably imitate the kind of generative humor capabilities humans possess.
Adam Krause. “Interstellar – TARS Humor Setting.” Online video clip. YouTube. Nov. 9, 2015.
Bar-Hillel, Yehoshua. Language and Information: Selected Essays on Their Theory and Application. Reading, MA, Addison-Wesley, 1964.
Beyond the Frame. “Louis CK and the Art of Non-Verbal Communication.” Online video clip. YouTube. Jun. 10, 2016. Web.
Berger, Arthur Asa. Blind Men and Elephants: Perspectives on Humor. New Brunswick, NJ, Transaction Publishers, 1995.
Berger, Arthur Asa. “Three Holy Men Get Haircuts: The Semiotic Analysis of a Joke.” Europe’s Journal of Psychology, vol. 12, no. 3, 2016, pp. 489–497. doi:10.5964/ejop.v12i3.1042.
Comedy Central. “Nathan for You – The Movement.” Online video clip. YouTube. Dec. 10, 2016. Web.
Dennett, Daniel C. “Cognitive Wheels: The Frame Problem of AI.” Minds, Machines and Evolution, 1984.
Dreyfus, Hubert L. “A History of First Step Fallacies.” Minds and Machines, vol. 22, no. 2, 2012, pp. 87–99. doi:10.1007/s11023-012-9276-0.
Galanter, Marc. Lowering the Bar: Lawyer Jokes and Legal Culture. Madison, WI, University of Wisconsin Press, 2005.
Haugeland, John. Having Thought: Essays in the Metaphysics of Mind. Cambridge, MA, Harvard University Press, 1998.
Haugeland, John. “Understanding Natural Language.” The Journal of Philosophy, vol. 76, no. 11, 1979, p. 619. doi:10.2307/2025695.
Irvine, Martine. “The Grammar of Meaning Systems: Sign Systems, Symbolic Cognition, and Semiotics.”
Jackendoff, Ray. Semantic Interpretation in Generative Grammar. Cambridge, MA, MIT Press, 1972.
Kuhlman, Thomas L. “A Study of Salience and Motivational Theories of Humor.” Journal of Personality and Social Psychology, vol. 49, no. 1, 1985, pp. 281–286. doi:10.1037//0022-3518.104.22.1681.
Larson, Erik J. “The Limits of Modern AI: A Story.” The Best Schools Magazine, www.thebestschools.org/magazine/limits-of-modern-ai/.
Manning, Christopher D., and Hinrich Schütze. Foundations of Statistical Natural Language Processing. Cambridge, MA, MIT Press, 2003.
Minsky, Marvin, and Seymour Papert. Perceptrons: An Introduction to Computational Geometry. Cambridge, MA, MIT Press, 1988.
Nijholt, Anton. “Incongruity Humor in Language and Beyond: From Bergson to Digitally Enhanced Worlds.” 14th International Symposium on Social Communication, 2015, pp. 594-599).
Norrick, Neal R. “Intertextuality in Humor.” Humor – International Journal of Humor Research, vol. 2, no. 2, 1989, doi:10.1515/humr.1922.214.171.124.
Schank, Roger C., and Robert P. Abelson. Scripts, Plans, and Knowledge. New Haven, Yale University, 1975.
Silver, Nate. The Signal and the Noise: Why so Many Predictions Fail—But Some Don’t. New York, NY, Penguin Press, 2012.
TED. “Heather Knight: Silicon-based comedy.” Online video clip. YouTube. Jan. 21, 2011. Web.
Ziv, Avner. Jewish Humor. New Brunswick, NJ, Transaction Publishers, 1998.