ABML: attention-based multi-task learning for jointly humor recognition and pun detection

Abstract

Humor and pun are widely used in our daily life. Due to the incongruous and inconsistent expressions in humor and pun, humor recognition and pun detection are challenging tasks in sentiment analysis and natural language processing. Pun is usually used as one of the humorous rhetorical devices. Thus, determining the puns in the texts can be pertinent to the successful recognition of humor and vice versa. Existing models of humor recognition and pun detection usually dealt with the two tasks independently. Therefore, we propose an attention-based multi-task learning model (ABML) to train the two tasks jointly. Our model unifies the two highly pertinent tasks, including the humor recognition and pun detection. In the ABML model, we design a Co-Encoder module to capture the common features between the two tasks by weight sharing. Apart from the Co-Encoder module, we also design two Private Encoder modules for the two tasks, respectively. The Private Encoder module is used to capture the private semantic feature of the two tasks. We conduct sufficient experiments on two widely used datasets, including a humor dataset and a pun dataset, to verify the effectiveness of our model. The experimental results show our model can get comparable performance compared with state-of-the-art methods.