XSYSIGMA at SemEval-2020 Task 7: Method for Predicting Headlines’ Humor Based on Auxiliary Sentences with EI-BERT

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 1077–1084 Barcelona, Spain (Online), December 12, 2020. 1077 XSYSIGMA at SemEval-2020 Task 7: Method for Predicting Headlines’ Humor Based on Auxiliary Sentences with EI-Bert Jian Ma , Shu-Yi Xie , Mei-Zhi Jin, Lian-Xin Jiang , Yang Mo , Jian-Ping Shen Ping An Life Insurance AI Department ShenZhen, China {MAJIAN446,XIESHUYI542,EX-JINMEIZHI001,JIANGLIANXIN769, MOYANG853,SHENJIANPING324}@pingan.com.cn Abstract This paper describes xsysigma team’s system for SemEval 2020 Task 7: Assessing the Funniness of Edited News Headlines. The target of this task is to assess the funniness changes of news headlines after minor editing and is divided into two subtasks: Subtask 1 is a regression task to detect the humor intensity of the sentence after editing; and Subtask 2 is a classification task to predict funnier of the two edited versions of an original headline. In this paper, we only report our implement of Subtask 2. We first construct sentence pairs with different features for Enhancement Inference Bert(EI-Bert)’s input. We then conduct data augmentation strategy and Pseudo-Label method. After that, we apply feature enhancement interaction on the encoding of each sentence for classification with EI-Bert. Finally, we apply weighted fusion algorithm to the logits results which obtained by different pre-trained models. We achieve 64.5% accuracy in subtask2 and rank the first and the fifth in dev and test dataset1, respectively. 1 Introduction Humor detection is a significant task in natural language processing. Most of the available humor datasets treat it as a binary classification task (Khodak et al., 2017; Davidov et al., 2010; Barbieri et al., 2014), i.e., examining whether a particular text is funny. It is interesting to study whether a short editing of a text can turn it from non-funny to funny. In this paper, we examine whether the headline is funniness or which word substitution is more funniness after a short editing. Such research helps us to focus on the humorous effect of word changes. Our goal is to determine how the machine understands the humor generated by such brief editing and to improve the accuracy of model prediction by different strategies, including rewriting the input sentences of model, enhancing local inference of features and applying the strategies of data augmentation and model ensemble. The competition dataset comes from Humicroedit (Hossain et al., 2019), a novel dataset for research in computational humor. Each edited text headline is labeled into a score of 0-3 by 5 judges with an average final score. The competition consists of two subtasks: Subtask 1 is to predict the average funniness of the edited headlines given the original headlines and the edited headlines and Subtask 2 is to predict which edited text is more funniness in comparison given the original headlines and two edited ones. This competition is significant because it is helpful for the task of generating humorous texts (Hossain et al., 2017), which can be applied to chatbots and news headline generation. In this paper, we only report our implementation on Subtask 2. We conduct several key data prepro- …

Content truncated to 500 words