Abstract: As a richly layered and complex linguistic expression,sarcasm is widely observed in people’s daily expressions and social platforms,and correctly detecting whether a comment has ironic intent in e-commerce,event topic analysis,etc.,is crucial to determine a commenter’s emotional tendency,attitude to the comment subject.Three types of contexts,namely,conversation context,user context and topic context,have been covered to build a context-rich sarcasm detection model.To address the problem that traditional shallow CNNs are difficult to capture sentence long-term dependencies,the proposed model introduces the DPCNN architecture to capture utterance remote association information and incorporates the bidirectional attention mechanism to learn incongruity information in conversation context.Considering the small number of sarcasm types and unbalanced levels of sarcasm expressions in realistic data samples,an asymmetric loss function with multiple learning modes is also proposed.Experiments are conducted on three public and real sarcasm datasets,and the results demonstrate that the method in this paper outperforms the benchmark model in ACC,F1 and AUC metrics by up to 2.5%,and the effectiveness of each module of the proposed model and the loss function of the multiple learning modes is demonstrated by ablation experiments,which can improve the performance of sarcasm detection.

Key words: Sarcasm detection, Context-rich, Bidirectional attention, Incongruity, Asymmetric loss