Fusion of Hand-crafted Features and Deep Semantic Features in a Unified Neural Model for Irony Detection in Microblogs

摘要

With the ever-growing popularity of microblogging platforms especially Twitter, people are increasingly interested in using these sites to convey their opinions, ideas, and beliefs. People often choose figurative language like irony or sarcasm to share their inherent emotions. Detecting and analyzing such ironic tweets might be beneficial for various natural language processing (NLP) tasks. Since tweet contents are noisy, identifying ironic tweets is a formidable task. In this paper, we present a neural model relying on high-level deep semantic features and hand-crafted features. To extract the higher-level deep features, we apply a combination of multi-kernel convolutional neural network (MKCNN) and long short-term memory (LSTM) network, where MKCNN is placed on top of the LSTM model. Besides, we exploit the lexical, sentimental, and Twitter-specific characteristics to extract a set of hand-crafted features and imputed them to a multilayer perceptron (MLP) module for better tweet representation. Feature representations from the MKCNN-LSTM module and MLP module are then combined and sent to the fully-connected irony detection module. Experiments on the SemEval-2018 irony detection dataset demonstrated the superiority of our presented model over most of the state-of-the-art methods.