PALI-NLP at SemEval-2022 Task 6: iSarcasmEval- Fine-tuning the Pre-trained Model for Detecting Intended Sarcasm

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 815 - 819 July 14-15, 2022 ©2022 Association for Computational Linguistics PALI-NLP at SemEval-2022 Task 6: iSarcasmEval- Fine-tuning the Pre-trained Model for Detecting Intended Sarcasm Xiyang Du and Dou Hu and Meizhi Jin and Lianxin Jiang and Yang Mo and Xiaofeng Shi Ping An Life Insurance Company of China, Ltd. {DUXIYANG037, HUDOU470, JINMEIZHI005 JIANGLIANXIN769, MOYANG853, SHIXIAOFENG309} @pingan.com.cn Abstract This paper describes the method we utilized in the SemEval-2022 Task 6 iSarcasmEval: Intended Sarcasm Detection In English and Arabic. Our system has achieved 1st in Sub- taskB, which is to identify the categories of intended sarcasm. The proposed system inte- grates multiple BERT-based, RoBERTa-based and BERTweet-based models with finetuning. In this task, our contribution is listed as follow: 1) we reveal several large pre-trained models’ performance on tasks coping with the tweet- like text. 2) Our methods prove that we can still achieve excellent results in this particular task without a complex classifier adopting some proper training method. 3) we found there is a hierarchical relationship of sarcasm types in this task. 1 Introduction Generally speaking, when we communicate through natural language, the literal meaning of the words is consistent with the meaning we want to express. Sarcasm is a form of linguistic expres- sion when this "congruence" is broken (Wilson, 2006). Due to the inherent metaphorical nature and sub- tle sentimental expression of this particular form of language expression. The detection task related to this kind of text, which is a negative expression of a positive emotion or the positive expression of neg- ative emotion, is extremely difficult for machines (Yaghoobian et al., 2021). This sarcasm data also weakens the detection modules that are widespread in our society (Maynard and Greenwood, 2014). Previous work shows that sarcasm often comes with incongruity between expectation and reality (Gibbs Jr et al., 1994). Many works attempt to model this incongruity within the text (Tay et al., 2018; Xiong et al., 2019). For multi-model data, some works use the features from different modali- ties (Schifanella et al., 2016; Cai et al., 2019), and some works shows that inter-modality incongruity is also an important feature for multi-modal sar- casm detection (Pan et al., 2020). The SemEval-2022 Task6 (Abu Farha et al., 2022) is designed to detect sarcasm and sarcasm types in twitter texts. In Subtask B, if a tweet is not sarcastic, it should not be annotated with any sarcasm label; if it is sarcastic, we need to detect which sarcasm it is, and it could have different sar- casm type at the same time. The main metric of this task is the Macro-F1 score of all sarcasm types. We design a simple and effective system for this task. The system is based on a large-scale pre- trained model based on bi-directional transformers (Vaswani et al., 2017) and fine-tuned to obtain the final output. First, we augmented the iSarcasm dataset with additional datasets, and then we set an appropriate learning rate for each layer …

Content truncated to 500 words