Ensemble-based domain adaptation on social media posts for irony detection

Social media provide platforms to express opinions on different issues or aspects like politics, products, brands, news events and so on. People often use irony while expressing their opinions. Verbal irony, an utterance that conveys a spirit completely opposed to the surface meaning expressed, is usually understood from the speaker’s body language and/or the context of the conversation. However, it is challenging to detect irony from a limited amount of written text automatically. In this paper, we study the issue of irony detection in social media posts collected during the 2019 general election in India. We use various machine learning and deep learning models (Bidirectional Encoder Representations from Transformers (BERT) and Embeddings fromLanguage Models (ELMo)) to classify them into irony and non-irony. We propose an ensemble model of machine learning and deep learning approaches. The classifiers are trained using a combined word embedding representation obtained from both BERT and ELMo.We create a dataset on the Indian General Election 2019 (IGE 2019 data) and then perform a series of experiments on irony detection, including a domain adaptation with the SemEval-2018 Task-3 (Sub-task A) dataset (SE-2018 T3 data). Our best performing model from machine learning based techniques as well as from deep learning based ones outperforms the state-of-the-art model in terms of accuracy, precision, recall, and F1 (metrics)