BERT’s The Word : Sarcasm Target Detection using BERT

BERT’s The Word: Sarcasm Target Detection using BERT Pradeesh Parameswaran, Andrew Trotman, Veronica Liesaputra, David Eyers Department of Computer Science University of Otago New Zealand [pradeesh,andrew,veronica,dme]@cs.otago.ac.nz Abstract In 2019, the Australasian Language Technol- ogy Association (ALTA) organised a shared task to detect the target of sarcastic com- ments posted on social media. However, there were no winners as it proved to be a difficult task. In this work, we revisit the task posted by ALTA using transformers— specifically BERT—given the current success of the transformer-based model in various NLP tasks. We conducted our experiments on two BERT models (TD-BERT and BERT- AEN). We evaluated our model on the data set provided by ALTA (‘Reddit’) and two addi- tional data sets: ‘book snippets’ and ‘Tweets’. Our results show that our proposed method achieves a 15.2% improvement from the cur- rent state-of-the-art system on the Reddit data set and a 4% improvement on Tweets. 1 Introduction Sarcasm is a remark made by a certain person to ridicule or hurt another person’s feelings (Cheang and Pell, 2008). A unique property of sarcasm lies in the way words are used. The result di- gresses from the conventional word order and al- ters the meaning of the whole sentence (Attardo et al., 2003). This very aspect also makes it very challenging to detect in a text. There has been a large number of studies that looked at automating sarcasm detection (Eke et al., 2020; Joshi et al., 2017) however there is less work done in identify- ing and extracting the target of sarcasm from the text. The problem of sarcasm target detection was originally coined by Joshi et al. (2016a). The tar- get of sarcasm is defined as an entity or a sit- uation that is being ridiculed in a sarcastic text. The task of sarcasm target identification is to ex- tract the subset of words that indicate the target of ridicule for a given sarcastic sentence. Iden- tifying the target of ridicule can improve the de- tection of cyber-bullying and hate speech targeted towards minority communities such as people of colour, the LGBTQ+ community and others (Oliva et al., 2021; Hylton, 2018). However, this task is particularly challenging because of the following factors: • Multiple targets—A sarcastic sentence may contain multiple targets. For instance in the following sentence, “James is as good at cooking as Guy Feiri is at avoiding contro- versy”, the targets are both “James” and “Guy Feiri”. • Lack of targets—The target of sarcasm may not be present in the given sentence. For ex- ample in the sentence, “I guess the kumara loves kayaking”, the speaker makes a sarcas- tic remark but the target of ridicule is unclear. When the sarcasm target does is not present or it is unclear, it is marked as OUTSIDE. There have been various attempts to improve the performance of sarcasm target detection (Pa- tro et al., 2019; Molla and Joshi, 2019; B¨ol¨uc¨u and Can, 2020; Parameswaran et al., 2021) such as through the use of …

Content truncated to 500 words