UoR-NCL at SemEval-2022 Task 6: Using ensemble loss with BERT for intended sarcasm detection

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 871 - 876 July 14-15, 2022 ©2022 Association for Computational Linguistics UoR-NCL at SemEval-2022 Task 6: Using ensemble loss with BERT for intended sarcasm detection Emmanuel Osei-Brefo University of Reading White Knights Campus Reading-UK e.osei-brefo@student.reading.ac.uk Huizhi Liang University of Newcastle Newcastle UK huizhi.liang@newcastle.ac.uk Abstract Sarcasm has gained notoriety for being difficult to detect by machine learning systems due to its figurative nature. In this paper, Bidirectional Encoder Representations from Transformers (BERT) model has been used with ensemble loss made of cross-entropy loss and negative log-likelihood loss to classify whether a given sentence is in English and Arabic tweets are sar- castic or not. From the results obtained in the experiments, our proposed BERT with ensem- ble loss achieved superior performance when applied to English and Arabic test datasets. For the validation dataset, our model performed better on the Arabic dataset but failed to out- perform the baseline method (made of BERT with only a single loss function) when applied on the English validation set. 1 Introduction In recent times, Social media platforms such as Twitter have turned out to be one of the most influential media for the expression of views, emotions, and information. To char- acterize sarcasm on Twitter; (Parmar et al., 2018) proposed the following: (a) strife be- tween negative situation and positive senti- ment, (b) strife between positive situation and negative sentiment, (c) Tweet starts with an interjection word, (d) likes and dislikes contra- diction, (e) tweet conflicting ubiquitous facts, (f) tweets with positive sentiment and antonym pair, and (g) tweet contrasting facts that are time-sensitive. With a huge amount of content being churned on social media and the need to ana- lyze and detect sarcasm closely, text classifica- tion methods have been widely introduced to deal with these sophisticated tasks. Sarcasm has been shown to pose a major difficulty for sentiment analysis models (Liu, 2010), mainly because sarcasm acts as a form of verbal irony which enables the concealment of the true in- tention of denigration and negativity under a pretence of open a respectful representation such as; “The only thing I got from college is a caffeine addiction". As a result, the ability to detect sarcasm is crucial for understanding the real intents and beliefs of people (Maynard and Greenwood, 2014). To address these issues and limitations, (Abu Farha et al., 2022) have organized the iS- arcasmEval shared task for intended Sarcasm Detection In English and Arabic where differ- ent tweets in English and Arabic languages have to be classified as either sarcastic or not. Several models for detection of sarcasm have been proposed in literature which incor- porate statistical, machine learning, and rule- based approaches but predominantly (Mandal and Mahto, 2019). However, these techniques are not able to effectively perceive the figura- tive and ironic meaning of words (Joshi et al., 2016). Also, these methods require manually engineered features and are unable to under- stand the patterns in passive voice sentences (Bajwa and …

Content truncated to 500 words