We present a novel multimodal preference dataset for creative tasks, consisting of over 250 million human ratings on more than 2.2 million captions, collected through crowdsourcing rating data for The New Yorker's weekly cartoon caption contest over the past eight years. This unique dataset supports the development and evaluation of multimodal large language models and preference-based fine-tuning algorithms for humorous caption generation. We propose novel benchmarks for judging the quality of model-generated captions, utilizing both GPT4 and human judgments to establish ranking-based evaluation strategies. Our experimental results highlight the limitations of current fine-tuning methods, such as RLHF and DPO, when applied to creative tasks. Furthermore, we demonstrate that even state-of-the-art models like GPT4 and Claude currently underperform top human contestants in generating humorous captions. As we conclude this extensive data collection effort, we release the entire preference dataset to the research community, fostering further advancements in AI humor generation and evaluation.

Large neural networks can now generate jokes, but do they really "understand" humor? We challenge AI models with three tasks derived from the New Yorker Cartoon Caption Contest: matching a joke to a cartoon, identifying a winning caption, and explaining why a winning caption is funny. These tasks encapsulate progressively more sophisticated aspects of "understanding" a cartoon; key elements are the complex, often surprising relationships between images and captions and the frequent inclusion of indirect and playful allusions to human experience and culture. We investigate both multimodal and language-only models: the former are challenged with the cartoon images directly, while the latter are given multifaceted descriptions of the visual scene to simulate human-level visual understanding. We find that both types of models struggle at all three tasks. For example, our best multimodal models fall 30 accuracy points behind human performance on the matching task, and, even when provided ground-truth visual scene descriptors, human-authored explanations are preferred head-to-head over the best machine-authored ones (few-shot GPT-4) in more than 2/3 of cases. We release models, code, leaderboard, and corpus, which includes newly-gathered annotations describing the image's locations/entities, what's unusual in the scene, and an explanation of the joke.

Humor is an integral aspect of the human experience. Motivated by the prospect of creating computational models of humor, we study the influence of the language of cartoon captions on the perceived humorousness of the cartoons. Our studies are based on a large corpus of crowdsourced cartoon captions that were submitted to a contest hosted by the New Yorker. Having access to thousands of captions submitted for the same image allows us to analyze the breadth of responses of people to the same visual stimulus.We first describe how we acquire judgments about the humorousness of different captions. Then, we detail the construction of a corpus where captions deemed funnier are paired with less-funny captions for the same cartoon. We analyze the caption pairs and find significant differences between the funnier and less-funny captions. Next, we build a classifier to identify funnier captions automatically. Given two captions and a cartoon, our classifier picks the funnier one 69% of the time for captions hinging on the same joke, and 64% of the time for any pair of captions. Finally, we use the classifier to find the best captions and study how its predictions could be used to significantly reduce the load on the cartoon contest's judges.

Obtaining useful crowdsourcing results often requires more responses than can be easily collected. Reducing the number of responses required can be done by adapting to previous responses with \textquotedbl{}adaptive\textquotedbl{} sampling algorithms, but these algorithms present a fundamental challenge when paired with crowdsourcing. At UW–Madison, we have built a powerful crowdsourcing data collection tool called NEXT (http://nextml.org) that can be used with arbitrary adaptive algorithms. Each week, our system is used by The New Yorker to run their Cartoon Caption contest (http://www.newyorker.com/cartoons/vote). In this paper, we will explain what NEXT is and it's applications, architecture and experimentalist use.

In this paper, we study the problem of automatically annotating the factoids present in collective discourse. Factoids are information units that are shared between instances of collective discourse and may have many different ways of being realized in words. Our approach divides this problem into two steps, using a graph-based approach for each step: (1) factoid discovery, finding groups of words that correspond to the same factoid, and (2) factoid assignment, using these groups of words to mark collective discourse units that contain the respective factoids. We study this on two novel data sets: the New Yorker caption contest data set, and the crossword clues data set.