“Judge me by my size (noun), do you?” YodaLib: A Demographic-Aware Humor Generation Framework

Proceedings of the 28th International Conference on Computational Linguistics, pages 2814–2825 Barcelona, Spain (Online), December 8-13, 2020 2814 “Judge me by my size (noun), do you?” YodaLib: A Demographic-Aware Humor Generation Framework Aparna Garimella♠†, Carmen Banea♠, Nabil Hossain‡ and Rada Mihalcea♠ ♠Computer Science and Engineering, University of Michigan †Adobe Research, India ‡Computer Science and Engineering, University of Rochester garimell@adobe.com, nhossain@cs.rochester.edu, {carmennb,mihalcea}@umich.edu Abstract The subjective nature of humor makes computerized humor generation a challenging task. We propose an automatic humor generation framework for filling the blanks in Mad Libs R⃝ stories, while accounting for the demographic backgrounds of the desired audience. We collect a dataset consisting of such stories, which are filled in and judged by carefully selected workers on Amazon Mechanical Turk. We build upon the BERT platform to predict location-biased word fillings in incomplete sentences, and we fine-tune BERT to classify location-specific humor in a sentence. We leverage these components to produce YODALIB, a fully-automated Mad Libs style humor generation framework, which selects and ranks appropriate candidate words and sentences in order to generate a coherent and funny story tailored to certain demographics. Our experimental results indicate that YODALIB outperforms a previous semi-automated approach proposed for this task, while also surpassing human annotators in both qualitative and quantitative analyses. 1 Introduction Computer-generated humor is an essential aspect in developing personable human-computer interactions. However, humor is subjective and can be interpreted in different ways by different people. Humor requires creativity, world knowledge, and cognitive mechanisms, which are extremely difficult to model theoretically. Generating humor is considered by some researchers as an AI-complete problem (Stock and Strapparava, 2002), hence humor generation was largely studied in specific settings. Language preferences vary with user demographics (Tresselt and Mayzner, 1964; Eckert and McConnell- Ginet, 2013; Garimella et al., 2016; Lin et al., 2018; Loveys et al., 2018), and this has led to approaches leveraging the demographic information of users to obtain better language representations and classi- fication performances for various NLP tasks (Volkova et al., 2013; Bamman et al., 2014; Hovy, 2015; Garimella et al., 2017). Humor is a universal phenomenon that is used across all countries, genders, and age groups (Apte, 1985). Likewise, there are variations in how humor is enacted and understood due to demographic differences (Kramarae, 1981; Duncan et al., 1990; Goodman, 1992; Alden et al., 1993; Hay, 2000; Robinson and Smith-Lovin, 2001). However, to the best of our knowledge, the effect of user demographics in computational humor generation has not been studied. Figure 1: Example of a partial Mad Lib story. In this paper, we introduce YODALIB, an au- tomatic humor generation framework for Mad Libs R⃝—a story-based fill-in the blank game— which also accounts for the demographic infor- mation of the audience and story coherence. We use location as the demographic dimension. YO- DALIB has three stages: (1) a candidate selection stage in which candidate words are selected to fill the sentences in each story, (2) a candidate ranking stage to assess if a filled-in (transformed) sentence is a …

Content truncated to 500 words