Year: 2021
Type: article
Abstract: Humor generation and classification are one the most challenging problems in computational Natural Language Understanding. Even humans fail at being funny and recognizing humor. This study attempts to create a joke generator using a large pre-trained language model (GPT2). Further, the authors develop a jokes classifier by fine-tuning pre-trained (BERT) to classify the generated jokes and attempt to understand what distinguishes joke sentence(s) from non-joke sentence(s). Qualitative analysis reveals that the classifier model has specific internal attention patterns while classifying joke sentences, which is absent when classifying normal sentences. The experimental results show the superiority of the BERT model compared to CNN and RNN+ attention baselines in terms of accuracy, precision, recall, and F1-score. The BERT model has achieved an accuracy of 0.983, precision (0.953), recall (0.978), and F1-score (0.964) Keywords: Deep Learning, Pre-trained, Joke Classifier, Generative Pre-trained Transformer 2 (GPT-2), Bidirectional Encoder Representations from Transformers (BERT).
Author: Irma Darmayanti Nur Arifin Akbar
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
Funder: Prince Sultan University
OpenAlex ID: W3200565935
