Classification and Regression Combined Model on Accessing Humor Score with Explanatory Feature
Yuchen Guo, Lingji Kong · 2022
Humor recognition and generation are still challenges for machine learning. There is a competition named Assessing Humor in Edited News Headlines. In the competition dataset, original headlines, edit words, and humor grades are provided. We utilized this dataset and tried to tackle the task, predicting the humor score of headlines after using the edit word to switch a word inside the original one. We divided the task into two major sections, the classification part and the regression part. Support vector machine (SVM) models are used for the classification task, and Bidirectional Encoder Representations from Transformers (BERT) models are used for the regression task. We proposed a new combined model by using classification labels to restrain the regression results. Through the task, we found that the models with a defining characteristic have a better performance. A further study on the explanatory models was conducted. Unfortunately, our classification model does not have an outstanding performance. Therefore, using classification labels to restrain the regression results does not get an excellent final model performance.