Human-written API documentation often becomes outdated, requiring developers to update it manually. Researchers have proposed identifying outdated API name references in documentation, yet have not addressed updating API documentation. Now, emerging large language models (LLMs) are capable of generating code examples and text descriptions. Then, a key question arises: Can LLMs assist in updating API documentation? In this paper, we propose an approach for leveraging an LLM to update API documentation with code change information. To evaluate this approach, we select five open-source projects that manage documentation revisions on GitHub and analyze the differences in documentation between two releases to derive ground truths. We then assess the accuracy of LLM-generated updates by comparing them to the ground truths. Our results show that LLM-generated updates achieve higher METEOR than outdated API documentation (0.771 vs 0.679). It indicates that the LLM updates are more similar to the human updates than the outdated documentation. Our results also reveal that LLMs update code-related information in API documentation with a maximum F1 score of$\mathbf{0. 9 2 1}$.

Prompt-based large language models (LLMs) have demonstrated their ability to perform tasks with minimal or no additional training data. In the context of issue classification, researchers have actively explored the capabilities of LLMs in classifying issue reports. However, existing studies still face limitations in accuracy. This study replicates an LLM-based issue classification study using GPT-3.5 Turbo and explores variants, such as adopting different models like Llama$3.18 B$and GPT-4o. Experimental results show that the classifier fine-tuned with GPT-3.5 Turbo still yields the same accuracy as shown in the original research and that the classifier fine-tuned with Llama$3.18 B(0.8004)$yields an F1-score of 0.0535 lower than that of the classifier fine-tuned with GPT-3.5 Turbo (0.8467). On the other hand, the classifier with GPT-4o (0.8639) yields an average$\mathbf{F 1}$-score$\mathbf{0. 0 1}$higher than that of the classifier fine-tuned with GPT-3.5 Turbo (0.8467). Additionally, the project-agnostic classifier fine-tuned with GPT-4o yields the highest$\mathbf{F 1}$-score of$\mathbf{0. 8 6 8 0}$. These findings contribute to advancing LLM-based issue classification by providing experimental insights into the accuracy of LLMs in this issue classification task.

In GitHub, open-source software (OSS) developers label issue reports. As issue labeling is a labor-intensive manual task, automatic approaches have developed to label issue reports. However, those approaches have shown limited performance. Therefore, it is necessary to analyze the performance of predicting labels for an issue report. Understanding labels with high performance and those with low performance can help improve the performance of automatic issue labeling tasks. In this paper, we investigate the performance of individual label prediction. Our investigation uncovers labels with high performance and those with low performance. Our results can help researchers to understand the different characteristics of labels and help developers to develop a unified approach that combines several effective approaches for different kinds of issues.