Humor in Collective Discourse: Unsupervised Funniness Detection in the New Yorker Cartoon Caption Contest

Humor in Collective Discourse: Unsupervised Funniness Detection in the New Yorker Cartoon Caption Contest Dragomir Radev1, Amanda Stent2, Joel Tetreault2, Aasish Pappu2 Aikaterini Iliakopoulou3, Agustin Chanfreau3, Paloma de Juan2, Jordi Vallmitjana2 Alejandro Jaimes2, Rahul Jha1, Robert Mankoff4 1 University of Michigan, 2 Yahoo! Research, 3 Columbia University, 4 The New Yorker radev@umich.edu, {stent,tetreaul,aasishkp}@yahoo-inc.com, {ai2315,ac3680}@columbia.edu {pdejuan,jvallmi,ajaimes}@yahoo-inc.com, rahuljha@umich.edu, bob mankoff@newyorker.com Abstract The New Yorker publishes a weekly captionless cartoon. More than 5,000 readers submit captions for it. The editors select three of them and ask the readers to pick the funniest one. We describe an experiment that compares a dozen automatic methods for selecting the funniest caption. We show that negative sentiment, human-centeredness, and lexical centrality most strongly match the funniest captions, followed by positive sentiment. These results are useful for understanding humor and also in the design of more engaging conversational agents in text and multimodal (vision+text) systems. As part of this work, a large set of cartoons and captions is being made available to the community. Keywords: collective discourse, computational humor, creativity 1. Introduction The New Yorker Cartoon Caption Contest has been running for more than 10 years. Each week, the editors post a car- toon (cf. Figures 1 and 2) and ask readers to come up with a funny caption for it. They pick the top 3 submitted cap- tions and ask the readers to pick the weekly winner. The contest has become a cultural phenomenon and has gener- ated a lot of discussion as to what makes a cartoon funny (at least, to the readers of the New Yorker). In this pa- per, we take a computational approach to studying the con- test to gain insights into what differentiates funny captions from the rest. We developed a set of unsupervised methods for ranking captions based on features such as originality, centrality, sentiment, concreteness, grammaticality, human- centeredness, etc. We used each of these methods to inde- pendently rank all captions from our corpus and selected the top captions for each method. Then, we performed Amazon Mechanical Turk experiments in which we asked Turkers to judge which of the selected captions is funnier. Figure 1: Cartoon number 31 Figure 2: Cartoon number 32 2. Related Work In early work, Mihalcea and Strapparava (2005) investi- gate whether classification techniques can distinguish be- tween humorous and non-humorous text. Training data consisted of humorous one-liners (15 words or less), and non-humorous one-liners, which are derived from Reuters news titles, proverbs, and sentences from the British Na- tional Corpus. They looked at features such as alliteration, antonymy and adult slang. Mihalcea and Pullman (2007) took this work further. They looked at four semantic classes relevant to human- centeredness: persons, social groups, social relationships, and personal pronouns. They showed that social relation- ships and personal pronouns have high prevalence in hu- mor. Mihalcea and Pullman also looked at sentiment; they found that humor tends to have a strong negative orienta- tion (especially in the case of long satirical text, but regular text also shows some …

Content truncated to 500 words