A Survey in Automatic Irony Processing: Linguistic, Cognitive, and Multi-X Perspectives

Proceedings of the 29th International Conference on Computational Linguistics, pages 824–836 October 12–17, 2022. 824 A Survey in Automatic Irony Processing: Linguistic, Cognitive, and Multi-X Perspectives Qingcheng Zeng Department of Linguistics Northwestern University qingchengzeng@outlook.com An-Ran Li Department of Chinese and Bilingual Studies The Hong Kong Polytechnic University an-ran.li@connect.polyu.hk Abstract Irony is a ubiquitous figurative language in daily communication. Previously, many re- searchers have approached irony from linguis- tic, cognitive science, and computational as- pects. Recently, some progress have been wit- nessed in automatic irony processing due to the rapid development in deep neural models in natural language processing (NLP). In this pa- per, we will provide a comprehensive overview of computational irony, insights from linguis- tic theory and cognitive science, as well as its interactions with downstream NLP tasks and newly proposed multi-X irony processing per- spectives. 1 Introduction Irony, which generally refers to the expressions that have opposite literal meanings to real meanings, is a representative rhetoric device in human languages (Li and Huang, 2020). For example, in sentences I just love when you test my patience!!, and Had no sleep and have got school now, compared to their literal meanings, both sentences are convey- ing reverse meanings and emotions, which mean not love and not happy. People are using ironies to communicate their affective states more implicitly or explicitly, to strengthen the claims depending on the needs, which contribute to neologism across human languages. However, the inner incongruity makes ironic expressions harder for machines to un- derstand. Consequently, accurate irony processing systems are essential and challenging for down- stream tasks and natural language understanding (NLU) research. Previously, many researchers approached irony processing from various perspectives. The last sur- vey in computational irony was published seven years ago (Wallace, 2015) and mainly focused on irony detection and pragmatic context model. Therefore, a systematic and comprehensive sur- vey on automatic irony processing remains ab- sent, encouraging us to focus on a review of ad- vances in irony processing, from traditional ma- chine learning, recurrent neural networks (RNNs) methods, to deeper pretrained language models (PLMs) throughout. Besides, del Pilar Salas-Zárate et al. (2020) revealed the great imbalance in figura- tive language research, in which sarcasm was domi- nant and twice as much as irony research. With this review, We aim to encourage a balanced and equal research environment in figurative languages. Moreover, the design and further improvement in present irony processing systems should be con- catenated with both the theoretical accounts in irony theories and its role in communication, fol- lowed by downstream applications in NLP tasks. Generally, most researchers approached irony pro- cessing as a single irony detection branch and three main shared tasks (Ghanem et al., 2019; Van Hee et al., 2018a; Ortega-Bueno et al., 2019) all focused on irony detection, whereas other aspects are con- stantly under-explored. And theoretical-informed or cognitive-informed discussions are rarely seen in irony research compared to tiny amendments in neural networks’ architectures. In this paper, we aim to offer a comprehensive review in irony …

Content truncated to 500 words