Recognizing Humour using Word Associations and Humour Anchor Extraction

Proceedings of the 27th International Conference on Computational Linguistics, pages 1849–1858 Santa Fe, New Mexico, USA, August 20-26, 2018. 1849 Recognizing Humour using Word Associations and Humour Anchor Extraction Andrew Cattle Xiaojuan Ma Hong Kong University of Science and Technology Department of Computer Science and Engineering Clear Water Bay, Hong Kong {acattle,mxj}@cse.ust.hk Abstract This paper attempts to marry the interpretability of statistical machine learning approaches with the more robust models of joke structure and joke semantics capable of being learned by neural models. Specifically, we explore the use of semantic relatedness features based on word asso- ciations, rather than the more common Word2Vec similarity, on a binary humour identification task and identify several factors that make word associations a better fit for humour. We also explore the effects of using joke structure, in the form of humour anchors (Yang et al., 2015), for improving the performance of semantic features and show that, while an intriguing idea, humour anchors contain several pitfalls that can hurt performance. 1 Introduction Humour is a part of everyday communication. Although telling and understanding jokes comes natu- rally to most humans, the recognition and interpretation of humour is a very difficult task for computers. Beyond merely expressing a funny idea, great jokes choose evocative words and present them in a sur- prising structure (Attardo, 2008) often with pleasing phonetics (Mihalcea and Strapparava, 2005). In addition to this linguistic mastery, many jokes require world knowledge either by setting up and then subverting an expectation or by referencing a common piece of culture (Kukovaˇcec et al., 2017). This makes computational humour a challenging yet very intriguing problem for natural language process- ing (NLP). As such, it is no surprise that computational humour, and humour recognition in particular, has received increased attention from the NLP community with SemEval-2017 devoting two tasks to it: ranking humorous tweets (Potash et al., 2017) and interpreting English puns (Miller et al., 2017). This recent attention has lead to advancements such as sequence-based neural humour models capa- ble of implicitly learning a joke structure and semantic features (Bertero and Fung, 2016a; Donahue et al., 2017). While these approaches offer good performance, their reliance on complicated neural ar- chitectures over explicitly engineered features present a problem for interpretability which may make it difficult to diagnose problems if results go wrong. Works which take a more interpretable statistical machine learning approach have their own draw- backs. For example, the representation of joke semantics has been fairly basic, typically computing word embedding similarities between all word pairs in a document (Yang et al., 2015), and bear little resemblance to the way humans actually interpret humour. Additionally, many works fail to take advan- tage of joke structure, treating texts as unordered bags-of-words (Bertero and Fung, 2016b; Mihalcea and Strapparava, 2005; Yan and Pedersen, 2017). This paper aims to marry the interpretability of statistical machine learning approaches with the more nuanced models of joke structure and joke semantics of neural approaches. Specifically, we explore the effectiveness modelling joke semantics …

Content truncated to 500 words