Detecting a homographic pun is one of the fundamental research tasks in natural language processing. A homographic pun is able to produce humor through the latent relationship between the pun and its semantically similar target. Puns have been widely applied in written and spoken forms of human language and have a long history. However, the ambiguity of a homographic pun is still a large challenge that cannot be addressed well with current methods. To alleviate this problem, we present a novel contextualized-representation gated attention (CRGA) network for the detection of homographic puns. This architecture has several advantages that can be exploited: one is that a contextual representation is used across varying linguistic contexts to address the polysemy of homographic puns; another other is that the CRGA model is able to detect homographic puns by combining the global semantic understanding, local script understanding, pun characteristic self-attention and gated mechanism. With this design, the CRGA model can effectively capture the polysemy information, which is helpful for homographic pun detection. The experimental results based on the common SemEval2017 Task7 and Pun of the Day datasets demonstrate the effectiveness and advancement of our proposed CRGA model.

检测同形双关语是自然语言处理中的基本研究任务之一。同形双关语能够通过双关语与其语义相似的目标之间的潜在关系产生幽默感。双关语已广泛用于人类语言的书面和口头形式，历史悠久。但是，同质双关语的含糊不清仍然是一个巨大的挑战，目前的方法无法很好地解决。为了减轻这个问题，我们提出了一种新颖的上下文表示门控注意（CRGA）网络，用于检测同构双关语。这种体系结构具有可以利用的几个优点：一个是在不同的语言上下文之间使用上下文表示来解决同构双关语的多义性。另一个是CRGA模型能够通过组合全局语义理解，局部脚本理解，双关特征自注意和门控机制来检测同形双关。通过这种设计，CRGA模型可以有效地捕获多义信息，这对单应双关语检测很有帮助。基于常见SemEval2017 Task7和Pun of the Day数据集的实验结果证明了我们提出的CRGA模型的有效性和先进性。