Artificial intelligence can do a lot of things today, but humor has not been its forte, at least until now. Researchers from the Hebrew University were able to teach a computer to identify articles that might have received an Ignoval Prize
In recent years, smart algorithms have been created in different and varied fields, which allow cars to move without a driver, clean the house (iRobot), draw pictures, compose music and even play computer games independently (remember Deep Blue?). However, teaching computers to understand, create and decode human humor, or even come close to it, still seem impossible tasks. A new study led by researchers from the Hebrew University - Prof. Dafna Shaf, Chen Shani and Nadav Bornstein from the School of Engineering and Computer Science and the Edmond and Lili Safra Neuroscience Center (ELSC) - sought to solve a sub-problem in the field of computational humor, and to examine whether computers can automatically recognize scientific achievements Funny, entertaining and unusual. The research will be presented in the framework ACL Conference on Natural Language Processing and Computational Linguistics Next month.
"In this work, we present a new task for detecting humor - identifying strange and funny scientific contributions. We draw inspiration from the 'Ig Nobel' prize, a satirical prize awarded annually to ten scientific achievements that 'first make people laugh and then make them think'. Past winners of the Ig Nobel included articles under the titles: 'Chickens prefer beautiful people' and 'Beauty is in the eyes of the beer holder: People who think they are drunk believe they are also attractive'. Ig Nobel Prize-winning essays provide a unique perspective on humor. On the one hand, humor is sophisticated, requires thought, as well as special knowledge and understanding of scientific culture. On the other hand, they can often be characterized as funny by the title alone, which is short, with simple syntax and no complex narrative structure (unlike longer jokes). That's why they are an interesting test case."
The researcher Chen Shani added in this context that "humor is an undefined and diverse field. Many times we can't even understand what makes something funny, or we don't agree on whether it's funny at all. Ig Nobel allows us to examine sophisticated and subtle (scientific) humor. Another reason to explore humor in science is that it is closely related to creativity and innovation - sometimes the oddball articles offer new perspectives that can lead to breakthroughs. For example, Andrei Geim won the Ig Nobel in Physics for using magnets to make frogs float. After a decade, he won the Nobel Prize for Physics, when the Nobel Prize Committee specifically attributed the prize to his playing with science."
The researchers created a first-of-its-kind database containing titles of funny scientific articles. They collected 211 articles that won the Ig Nobel and another 1,496 humorous articles that they found on the Internet (mostly in online forums and blogs). In addition, examples of non-funny articles were also sampled (a random sample of 1,707 different publications). They then classified each article into one of the following scientific fields: neuroscience, medicine, biology, or exact sciences. After that, the researchers built a model that connects machine learning and insights from humor research literature (psychology, linguistics, etc.).
After reviewing the literature, the researchers chose to focus on four main elements: (1) the element of surprise - the researchers trained language models to generate for the computer expectations about normal subjects for scientific research; (2) Simplicity - the researchers stated in their article that "We hypothesize that funny article titles tend to be simpler", so title lengths and word lengths were calculated. Also, the researchers used different measures of language difficulty, for example the age at which children acquire a certain word ("dog" vs. "photosynthesis"); (3) foul language - the algorithm tries to predict the degree of bluntness of the research; (4) Funny language - because the use of funny words can imply that the article containing them is entertaining. This variable is tested by several indicators, one of which is an algorithm that the researchers trained to identify short jokes (one-liners).
At the end of the process, the algorithmic model received as input titles of scientific articles and outputted a binary score (funny/serious), and a level of confidence in this score. According to the researchers, according to the confidence level of the model, millions of articles can be sorted according to their "funny" level, "that is, how funny our model thinks the research described in the article is. Given the title of an article, we can assess, through the algorithm, if it is funny and with what probability," the researchers explain. To evaluate the algorithm, the researchers conducted an experiment outside of their original database, identifying funny content within a pool of over 0.6 million articles. In other words, their algorithmic model was able to flood funny articles as relevant candidates for the Hague Nobel Prize.
The difficulty in creating the algorithm was in "translating" non-computational literature, which comes from fields such as psychology, philosophy and linguistics, into defined measures that can be identified with the help of a computer. For example, creating expectations required a lot of creativity, since these are things that are very clear to us as humans (and researchers), but very non-trivial for a computer. Despite the difficulties, the results of the research were impressive. Examples that the algorithm recommended were, for example: an article from neuroscience about whether we bring the food closer to the mouth or the mouth to the food; An article from a monkey study that tests whether Nancy the chimpanzee says "no" by shaking her head; and an article from psychology that investigated why and in what situations people lie on the Internet (for example, more on dating sites than on social networks). The most entertaining studies, in case you were wondering, came from the field of psychology.
Another surprise of the researchers was that more funny articles were found in social sciences and medicine compared to exact sciences. "We notice that most of the articles classified by our model as funny belong to the social sciences ("Dogs can distinguish smiling human faces through casual expressions") or medicine ("Can monkeys tell us about human amnesia when in fact they can't talk at all?"), compared For exact sciences ("Kinematics of eating with a spoon - bring the food to the mouth, or the mouth to the food?"). We believe that this happens because often the social sciences and the world of medicine present topics that are more familiar to laymen (the titles are written accordingly)," the article reads.
Prof. Dafna Shaf and Chen Shani concluded: "Humor is a very complicated and diverse phenomenon, depending on the context, person and culture. Computers that know how to deal with humor in all its shades is a matter that currently belongs to the worlds of science fiction, but they are now able to deal quite well with specific tasks, such as the one we presented in the current study. In addition to our original goal (finding candidates for the Ig Nobel), the ability to identify entertaining articles gives us a new and interesting angle to the questions of Science of Science. For example, at what stage of their career do researchers tend to conduct more unusual and entertaining research, does such research receive more exposure, and more."
