Ferryman at SemEval-2020 Task 7: Ensemble Model for Assessing Humor in Edited News Headlines

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 1008–1012 Barcelona, Spain (Online), December 12, 2020. 1008 Ferryman at SemEval-2020 Task 7: Ensemble Model for Assessing Humor in Edited News Headlines Weilong Chen, Jipeng Li, Chenghao Huang, Wei Bai, Yanru Zhang, and Yan Wang∗ University of Electronic Science and Technology of China chenweilong1995,jeeperly,huangchenghao,cellur@std.uestc.edu.cn yanruzhang,yanbo1990@uestc.edu.cn Abstract Natural language processing (NLP) has been applied to various fields including text classification and sentiment analysis. In the shared task of assessing the funniness of edited news headlines, which is a part of the SemEval 2020 competition, we preprocess datasets by replacing abbrevi- ation, stemming words, then merge three models including Light Gradient Boosting Machine (LightGBM), Long Short-Term Memory (LSTM), and Bidirectional Encoder Representation from Transformer (BERT) by taking the average to perform the best. Our team Ferryman wins the 9th place in Sub-task 1 of Task 7 - Regression. 1 Introduction With the increased influence of social media on modern society, massive news emerges on the Internet every single day. To attract people’s attention, editing news headlines to make them more humorous is now an effective approach. However, considering the amount of headlines, it is impractical to manually modify each headline one by one. To solve this problem, some automatic methods to add humor in news headlines are needed. Assessing the funniness of edited news headlines plays a significant role there. Here we present the method to automatically assessing the funniness of edited news headlines using machine learning algorithms. In SemEval-2020 Task 7: Assessing the Funniness of Edited News Headlines, nearly all existing humor datasets are annotated to study whether a short text is funny(Hossain et al., 2020). The goal of the task is to determine how machines can understand humor generated by short edits. The task has been divided into two sub-tasks: 1) given the original and edited headline, predicting the mean funniness of the edited headline; 2) given the original headline and two edited versions, predicting which edited version is the funnier of the two. The two sub-tasks are independently evaluated by root mean square error (RMSE) on the overall test set. Since our team has participated in multiple tasks in SemEval-2020 simultaneously, we only selected the Sub-task 1 in Task 7. In this task, we encountered two main problems: (i) Training single model has no significant effect; (ii) The raw dataset fits Bidirectional Encoder Representation from Transformer (BERT) badly. To address problem (i), we adapt and fine-tune the merged model, such as increasing weights of edited words in Light Gradient Boosting Machine (LightGBM), and input the sentences before and after modification as features in Long Short-Term Memory (LSTM). Then we merge three models, which are LightGBM, LSTM and BERT respectively. To accelerate model training, we improve LightGBM with Word2Vec, FastText, and Global Vectors for Word Representation (GloVe). And to reduce , we use LSTM with max pooling and min pooling, supplementing with Enhanced Sequential Inference Model (ESIM). We also use multi-sample dropout, Frequent Pattern Growth (FPG), 5-fold cross validation, and …

Content truncated to 500 words