A Modular Architecture for Unsupervised Sarcasm Generation

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 6144–6154, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 6144 A Modular Architecture for Unsupervised Sarcasm Generation Abhijit Mishra, Tarun Tater, Karthik Sankaranarayanan IBM Research {abhijimi,ttater24,kartsank}@in.ibm.com Abstract In this paper, we propose a novel framework for sarcasm generation; the system takes a lit- eral negative opinion as input and translates it into a sarcastic version. Our framework does not require any paired data for training. Sarcasm emanates from context-incongruity which becomes apparent as the sentence un- folds. Our framework introduces incongruity into the literal input version through modules that: (a) filter factual content from the in- put opinion, (b) retrieve incongruous phrases related to the filtered facts and (c) synthe- size sarcastic text from the filtered and incon- gruous phrases. The framework employs re- inforced neural sequence to sequence learn- ing and information retrieval and is trained only using unlabeled non-sarcastic and sarcas- tic opinions. Since no labeled dataset exists for such a task, for evaluation, we manually prepare a benchmark dataset containing lit- eral opinions and their sarcastic paraphrases. Qualitative and quantitative performance anal- yses on the data reveal our system’s superior- ity over baselines, built using known unsuper- vised statistical and neural machine translation and style transfer techniques. 1 Introduction Sarcasm1 is an intensive, ironic construct that is intended to express contempt or ridicule. It is of- ten linked with intelligence, creativity, and wit, and therefore empowering machines to generate sarcasm is in line with the key goals of Strong AI2. From the perspective of Natural Language Generation (NLG), sarcasm generation remains an important problem and can prove useful in downstream applications such as conversation sys- tems, recommenders, and online content genera- tors. For instance, in a conversational setting, a 1https://www.thefreedictionary.com/Sarcasm 2https://en.wikipedia.org/wiki/Artificial general intelligence more natural and intriguing form of conversation between humans and machines could happen if machines can intermittently generate sarcastic re- sponses, like their human counterparts. Over the years, a lot of research and develop- ment efforts have gone into the problem of detect- ing sarcasm in text, which aims to classify whether a given text contains sarcasm or not (Joshi et al. (2017b) provide an overview). However, systems for generation of sarcasm have been elusive. This is probably due to the fact that in sarcasm gener- ation both selection of contents for sarcastic opin- ion generation and surface realization of contents in natural language form are highly nuanced. In the broader area of style transformation of texts, most of the existing works have focused narrowly on transformations at lexical and syn- tax levels, i.e., text simplification (Siddharthan, 2014), text formalization (Jain et al., 2018), sen- timent style transfer (Shen et al., 2017; Xu et al., 2018), sentiment flipping (Li et al., 2018) and understanding humor (West and Horvitz, 2019). However, very little work has been done ((Piwek, 2003),(Hovy, 1987)) on incorporating pragmat- ics into generation tasks such as …

Content truncated to 500 words