One does not simply produce funny memes! - Explorations on the Automatic Generation of Internet humor

This paper reports on the automatic generation of image macro based Internet memes – potentially funny combinations of an image and text, intended to be spread in social networks. Memes are produced for news headlines for which, based on linguistic features, a suitable macro is selected and the text is adapted. The generation method is described together with its current implementation, which integrates a variety of tools and resources. Illustrative examples are also presented. Results of a human evaluation showed that, despite positive and neutral assessments, overall, automaticallygenerated memes are still below those produced by humans. Introduction Internet memes are a current trend, typically jokes, that gain influence through transmission in social media (Davison, 2012). A popular kind is the image macro – a set of stylistic rules for adding text (e.g. “One does not simply X”, “What if I told you Y”) to images, with a specific semantics. Most memes are a product of human creativity and their automatic generation is thus a challenge for computational creativity. This paper reports on an exploratory approach for the automatic generation of Internet memes for news headlines – or better, protomemes, which may eventually become memes if spread through the Web. Following our previous work (Costa, Gonçalo Oliveira, and Pinto, 2015), where memes were generated for trendy people, towards whom famous quotes were adapted and added to their images, we now present a system with a similar goal, but with significant differences. Given a headline, MEMEGERA 2.0: automatically selects a suitable well-known image macro from a predefined set; adapts the text according to the macro rules; and combines the text with the image. The result is ready to be consumed. Following the recent trend of using Twitter as a showcase for linguistic creativity (Veale, Valitutti, and Li, 2015), an implemented Twitterbot posts a new meme every hour. Besides the main challenge, the bot’s feed can be used as an alternative and funny way of following recent news. The illustrative examples presented here, as well as the memes posted on Twitter, confirm that our main goal was achieved: given a headline, a coherent combination of image and text, easily recognisable as an Internet meme, and with a relation to the headline, is produced. A more challenging goal, discussed in the end of the paper, involves the production of novel artefacts with humor value. Though limited in terms of size, an evaluation survey pointed out that, despite frequently producing coherent text, there is work to do in the selection of the suitable image macros, as well as on the surprise and humor value of the produced artefacts. Yet, although better than MEMEGERA 2.0 overall, human-generated memes also failed to consistently reach top-performance, which shows that the creation of memes is challenging, even for humans. In the remainder of this paper, we provide background knowledge and work on related topics, including computational linguistic creativity and computational humor. The meme generation method is then described and followed by details on its current implementation, for Portuguese, our native language. MEMEGERA 2.0 exploits a variety of tools and resources for collecting data, processing text, and for combining and publishing the results. Several illustrative memes generated by MEMEGERA 2.0 are then shown and contextualised. Before concluding, the conducted evaluation and its results are discussed. Background and Related Work Given a news headline, MEMEGERA 2.0 generates Internet memes that combine an image and a piece of text. This involves the automatic selection of an adequate image macro for the headline and the adaptation of the headline text according to the macro, with a humor intent. After a short introduction on the concept of Internet memes, this section enumerates previous work on computational linguistic creativity with a focus on humor generation. Not forgetting the role the image plays on the memes, the section ends with a reference to visual humor and its combination with text. Internet Memes The term meme originally refers to an idea, behaviour, or style that spreads from person to person within a culture (Dawkins, 1976). In the Web 2.0 era, it was adopted to denote a piece of culture, typically a joke, which gains influence through online transmission (Davison, 2012). A popular kind of meme is the image macro, which involves a set of stylistic rules for adding text to images. The same text can be added to different images or different text can be added to a common image. We focus on latter. Yet, the new piece of text should be analogous to the original, which is