Sarcasm is a form of figurative language where the literal meaning of words cannot hold, and instead the opposite interpretation is intended in a text. Sarcasm detection is a significant task to mine fine-grained information, which is a much more difficult challenge for sentiment analysis. Both industry and academia have realized the importance of sarcasm detection. However, most existing methods do not work very well. Using a neural architecture, we propose a novel multi-dimension question answering (MQA) network in order to detect sarcasm. MQA not only introduces the abundant semantic information to understand the ambiguity of sarcasm by multi-dimension representations, but also builds the conversation context information by deep memory question answering network based on bidirectional LSTM and attention mechanism to discover sarcasm. The experimental results show that our model has ability to obviously outperform other state-of-the-art methods, and then further examples also verifies the advancement and effectiveness of our proposed network for detecting sarcasm.

In recent years, hate speech spread on the Internet has seriously affected society’s harmony, stability, and development. A way to quickly identify hate speech from the vast amount of data on the Internet is urgent. In this paper, different from previous traditional methods, we explore a novel scenario of constructing a syntax dependency graph for each instance based on the syntactical information retrieved from an external tool. We propose a model called the Dependency Graph Convolutional and Sentiment Knowledge Transfer (DGCSKT) network. DGCSKT utilizes syntactic dependency graphs and dependency graph convolutional operations to enhance the model’s ability to perceive contextual information. Additionally, we introduce sentiment resources that are data-homogeneous as an auxiliary task at the bottom level of the model to share effective sentiment features and improve recognition performance. Then, we propose the Dynamic Normalized Weighting (DNW) method to weight the training information of different tasks and thus improve the model’s generalization ability. Compared to the current state-of-the-art methods, our proposed approach improves the Macro-F1 by 3.88% and 0.54% in OLID and HateEval respectively.