Humor is an inherently social phenomenon, with humorous utterances shaped by what is socially and culturally accepted.

Understanding humor is an important natural language processing challenge, with many applications to human-computer interactions. From conversations with Amazon’s Alexa to predicting party game winners and comic caption winners, to automating online search functionality in a gamified way, Dafna Shahaf is exploring humor in the context of everyday occurrences alongside research partners from Amazon, Yahoo, Microsoft, and many other institutions worldwide. Highlights of Dafna’s work to humanize computers through humor include:

• Characterizing Playful Requests to Conversational Agents – “Alexa, do you want to build a snowman?” Other Conversational Agents (CAs), such as Apple’s Siri and Amazon’s Alexa, are well-suited for task-oriented interactions (“Call Jason”), but other interaction types are often beyond their capabilities. One notable example is playful requests: for example, people ask their CAs personal questions (“What’s your favorite color?”) or joke with them (“Find Nemo”). Failing to recognize playfulness causes user dissatisfaction and abandonment, destroying the precious rapport with the CA. Today, playful CA behavior is achieved through manually curated replies to hard-coded questions. Prof. Shahaf’s research takes a step toward understanding and scaling playfulness by characterizing playful opportunities. To map the problem’s landscape, she draws inspiration from humor theories and analyzes real user data. Her team presents a taxonomy of playful requests and explores their prevalence in actual Alexa traffic to inspire new paths to more human-like CAs.
• Identifying Humorous Cartoon Captions – Motivated by the prospect of creating computational models of humor, Shahaf’s team studies the influence of the language of cartoon captions on the perceived humorousness of the cartoons. 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 the team to analyze the breadth of responses of people to the same visual stimulus.
• Predicting Humor in a Fill-in-the-Blank Party Game – In this work, Shahaf’s team explores humor in the context of Cards Against Humanity – a party game where players complete fill-in-the-blank statements using cards that can be offensive or politically incorrect. They introduced a novel dataset of 300,000 online games of Cards Against Humanity, including 785K unique jokes, analyzed it, and provided insights. They then trained machine learning models to predict the winning joke per game, achieving performance twice as good vs.  random, even without any user information. Researchers found that the models are primarily focused on punchline cards, with the context having little impact. Analyzing feature importance, they observed that short, crude, juvenile punchlines tend to win.