An attention network via pronunciation, lexicon and syntax for humor recognition

摘要

Humor is one of the most common and attractive expressions in our daily life. It is usually witty and funny. Humor recognition is an interesting but difficult task in natural language processing. Some recent works have used deep neural networks to recognize humorous text. In a different approach, we start from a new perspective based on humor linguistics, including pronunciation, lexicon, and syntax, for recognizing humor based on neural networks, in order to capture humorous incongruity and ambiguity. Specifically, we propose an attention network via pronunciation, lexicon, and syntax (ANPLS) for humor recognition. The ANPLS model contains four units, namely, the pronunciation understanding unit, the lexicon understanding unit, the syntax analysis unit, and the context understanding unit. The pronunciation understanding unit is used to extract the pronunciation-based humor features. The lexicon understanding unit is used to solve the polysemy in humor. The syntax analysis unit aims to capture the syntax information of humor. The context understanding unit is used to obtain the contextual humor features. These four units may have different levels of importance for humor recognition so that we further apply an attention mechanism to assign different weights to these four units. We conduct experiments on three popular datasets, namely, the SemEval2017 Task7 dataset, the 16000 One-Liners dataset, and the Pun of the Day dataset. The experimental results demonstrate that our model can achieve comparable or state-of-the-art performance compared with the existing models.