Sarcasm detection based on BERT and attention mechanism

Sarcasm detection is a challenging task in sentiment analysis and is usually used to detect sarcasm by judging inconsistencies in the individual words of an expression of sentiment; however, detection is less effective for sentences with complex semantic information, especially those with too few sentiment words. To tackle this problem, we propose a sarcasm detection model based on pretraining model and attention mechanism, which works with the context and language environment for sarcasm detection of phrase fragments for semantic feature extraction to attain higher-level semantic features, rather than limiting attention to words. Intrasentence attention mechanism was used to model the semantic information between phrase fragments within sentences, and then further strengthen the key features, which gives higher attention weights to key phrases so as to improving the ability of the model to identify expressions of sarcasm and avoids the lack of semantic incongruity of distant words in most sequential neural networks. Experimental results on a publicly available dataset indicated that the proposed approach outperforms baselines and state-of-the-art models.