Keskeiset käsitteet
State-of-the-art language models can generate humorous allegorical sayings in Chinese, but there is room for improvement in matching human creativity.
Tiivistelmä
This research delves into the generation of Chinese humor through two training methods: fine-tuning a medium-sized language model and prompting a large one. The study focuses on two-part allegorical sayings, a unique form of Chinese humor rich in wordplay and cultural insights. The paper explores the challenges of computational understanding and generation of humor, particularly in non-English languages like Chinese. It investigates the effectiveness of state-of-the-art language models in comprehending and generating Chinese humor, emphasizing the importance of training paradigms to enhance humor elements. Results show that while these models can generate humorous allegorical sayings, there is still a noticeable gap in quality compared to human-created content.
Tilastot
Human-annotated results show that models can generate humorous allegorical sayings.
Fine-tuning LM achieved coherency/humor scores of 2.13/1.59.
ChatGPT model with few-shot prompting obtained scores of 2.45/1.32.
Gold human-written samples received the highest scores of 2.89/2.06.