Knowledge prompt turning (KPT) is a prompt learning method, which aims to improve the classification performance of deep learning models by optimizing the mapping from predicted label words to categories under the condition of using natural language prompt words. However, this cloze-style approach that only tunes word mapping neglects the importance of natural language prompt stems, and its existing application fields are limited to natural language processing. This study proposes a new method to optimize the natural language prompt stem in KPT, using a prompt word optimization method based on Monte Carlo search trees, guiding large language model (LLM) to generate optimal prompt words for task models, thereby improving the model’s performance in software defect detection tasks using KPT. This study demonstrates the significant improvement of optimizing prompt stems on the performance of deep learning models and verifies the effectiveness of this new type of prompt learning method in software defect detection tasks.

In collaborative software development, bots have become increasingly prevalent, making effective bot recommendation a key factor in enhancing development efficiency. This study aims to explore the application and efficacy of deep learning models in bot recommendation. Focusing on the CodeBERT model, we conduct a comprehensive evaluation through comparison with baseline models, parameter tuning (including batch size and learning rate), and the incorporation of language data. Our findings demonstrate that under specific conditions, the CodeBERT model exhibits superior performance in bot recommendation tasks, with parameter adjustments and the inclusion of language data significantly impacting the model's effectiveness. These insights offer new perspectives and strategies for the effective recommendation of bots in open-source software platforms.

Bots are prevalent contributors in collaborative software development, necessitating accurate detection techniques. This preliminary study aims to leveraging public datasets and Transformer-based models (i.e., BERT, CodeBERT, RoBERTa, BART, and PLBART) for the bot detection task. Our experimental result reveals that CodeBERT achieves the highest performance, with an impressive accuracy score of 94.1%.