Enhancing Novelty in ChatGPT Responses: Incorporating Random Word Brainstorming
Pittawat Taveekitworachai, Ruck Thawonmas · 2023
This paper presents a new prompting approach for increasing the novelty in ChatGPT responses. ChatGPT has proven to be effective in generating natural language responses; however, ensuring response novelty remains challenging. Our proposed method, inspired by random word brainstorming, includes random words in prompts to introduce more diversity in ChatGPT responses. Through a questionnaire-based evaluation, we compared preferences for solution ideas generated using the standard approach and our proposed approach. We found that participants preferred our technique in 65% of the 20 problems. The results suggest the effectiveness of our proposed approach. We also explored the use of GPT models as evaluators, with GPT-3.5 achieving 65% accuracy and GPT-4 achieving 70% accuracy when compared to human preferences from the questionnaire. These results suggest the potential of leveraging GPT models as noisy natural language evaluators. For future studies, we recommend focusing on prompt engineering and word list design to further improve performance. Overall, incorporating random words in prompts can effectively increase novelty in ChatGPT responses.