X-PuDu at SemEval-2022 Task 6: Multilingual Learning for English and Arabic Sarcasm Detection

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 999 - 1004 July 14-15, 2022 ©2022 Association for Computational Linguistics X-PuDu at SemEval-2022 Task 6: Multilingual Learning for English and Arabic Sarcasm Detection Yaqian Han, Yekun Chai, Shuohuan Wang, Yu Sun Baidu, Inc. {hanyaqian,chaiyekun,wangshuohuan,sunyu02}@baidu.com Hongyi Huang, Guanghao Chen, Yitong Xu, Yang Yang Shanghai Pudong Development Bank {huanghy6,chengh13,xuyt3,yangy103}@spdb.com.cn Abstract Detecting sarcasm and verbal irony from peo- ple’s subjective statements is crucial to under- standing their intended meanings and real sen- timents and positions in social scenarios. This paper describes the X-PuDu system that partic- ipated in SemEval-2022 Task 6, iSarcasmEval - Intended Sarcasm Detection in English and Arabic, which aims at detecting intended sar- casm in various settings of natural language un- derstanding. Our solution finetunes pre-trained language models, such as ERNIE-M and De- BERTa, under the multilingual settings to rec- ognize the irony from Arabic and English texts. Our system ranked second out of 43, and ninth out of 32 in Task A: one-sentence detection in English and Arabic; fifth out of 22 in Task B: binary multi-label classification in English; first out of 16, and fifth out of 13 in Task C: sentence-pair detection in English and Arabic. 1 Introduction Sarcasm is the use of language that typically signi- fies the opposite to mock or convey contempt. As a narrow research field in natural language process- ing (NLP), sarcasm detection is a particular case in the spectrum of sentiment analysis, with important implications for a slew of NLP tasks, such as sen- timent analysis, opinion mining, author profiling, and harassment detection. In the textual data, these tonal and gestural clues like heaving tonal stress and rolling of the eyes are missing, making it more difficult for machines. The sarcastic intention of human annotators has potentially hindered the training and evalu- ation process in detecting the genuine emotions and positions of the natural language. Thus, this task (Abu Farha et al., 2022) adopted a novel data collection method (Oprea and Magdy, 2020), where authors themselves label the training sam- ples. For sarcastic texts, the authors also rephrase them into non-sarcastic ones. Then, linguistic ex- perts further checked the scathing pieces and la- beled them into sub-categories of sarcasm defined by (Leggitt and Gibbs, 2000): sarcasm, irony, satire, understatement, overstatement, and rhetorical ques- tion. This SemEval task requires the identification of sarcasm in either one sentence or sentence pairs in various language settings, which consists of three subtasks: (1) Task A (English and Arabic): Given a text, determine whether it is sarcastic or non- sarcastic; (2) Task B (English only): A binary multi-label classification task. Given a text, determine which ironic speech category it belongs to, if any; (3) Task C (English and Arabic): Given a sarcas- tic text and its non-sarcastic rephrase, i.e. two texts that convey the same meaning, determine which is the sarcastic one. Our method employed various multilingual or mono-lingual pre-trained language models, such as ERNIE-M (Ouyang et al., 2020) and DeBERTa (He et al., …

Content truncated to 500 words