ColBERT: Using BERT Sentence Embedding for Humor Detection

Automatic humor detection has interesting use cases in modern technologies, such as chatbots and virtual assistants. Based on the general linguistic structure of humor, in this paper, we propose a novel approach for detecting humor in short texts by using BERT sentence embedding. Our proposed method uses BERT to generate embeddings for sentences of a given text and uses these embeddings as inputs for parallel lines of hidden layers in a neural network. These lines are finally concatenated to predict the target value. For evaluation purposes, we created a new dataset for humor detection consisting of 200k formal short texts (100k positive and 100k negative). Experimental results show that our proposed method can determine humor in short texts with accuracy and an F1-score of 98.2 percent. Our 8-layer model with 110M parameters outperforms all baseline models with a large margin, showing the importance of utilizing linguistic structure in machine learning models.