評估ChatGPT-4o 的R 語言程式碼生成共字網絡圖的效能：人－機－人的協同模式

Evaluating the Performance of ChatGPT-4o in Generating Co-word Network Diagrams with R Programming: Establishing a Human-AI-Human Collaboration Model

隨著對話式大型語言模型（簡稱AI）的興起，其在電腦程式碼生成與資料視覺化領域的應用日益廣泛。然而，AI自動生成R語言共字網絡圏腳本的能力仍有待系統性驗證。因此，本研究旨在評估ChatGPT-4o於生成各類共字視覺化圖形的R程式碼效能，並確立人－機－人作模式的可行性。
本研究依循人－機－人流程，開發10組「處理端」R語言範本（題庫），對應「輸入端」資料（項目實體與關聯矩陣）繪製「輸出端」的共字網絡圖。評估時的測驗（題目），係將輸入資料與預期圖例提問ChatGPT-4o生成R碼腳本程式。評量方法（計分），則以製圖符合答案與再提問互動次數（每追加提示一次扣1分）進行成效的判準，最後將得分換算為百分比以代表AI整體表現。
結果顯示，ChatGPT-4o在10組測試中達成90％之正確率，能於大多數情境下準確解析資料並產生有效的R程式碼。惟於較複雜圖形或具方向性要求時，仍須進行人為微調與修正。
本研究驗證了AI於「輸入端」資料處理與「處理端」程式碼生成的高效性，同時突顯了「輸出端」成果仍需專業人工最終確認的重要性。人－機－人協作模式展現出未來於大規模文本探勘與知識網絡建構應用中的高度實用潛力。

With the rise of conversational large language models (LLMs), their applications in computer code generation and data visualization have expanded rapidly. However, the ability of AI to generate R scripts for co-word network visualization remains to be systematically verified. This study aimed to evaluate the performance of ChatGPT-4o in generating R scripts for various types of visualizations and to establish the feasibility of a human-AI-human collaborative model.
Following a human-AI-human workflow, we first developed ten R script templates (''processing modules'') capable of generating co-word network visualizations (''output'') based on provided input data (''input modules'' consisting of entity and relationship files, serving as item banks). We then submitted the input data and corresponding target figures to ChatGPT-4o to generate R scripts as test items. Performance was evaluated based on the number of additional prompts required: each additional prompt resulted in a one-point deduction, and the final scores were converted into percentages to represent the overall performance of the AI under the established scoring criteria.
Results showed that ChatGPT-4o achieved a 90% overall success rate across the ten test cases, correctly interpreting the input data and producing effective R visualization scripts in most scenarios. However, for more complex or directionally structured figures, human intervention remained necessary for final refinements.
This study confirms that AI can efficiently handle input data processing and rapidly generate preliminary R scripts. Nevertheless, the output still requires human verification, underscoring the necessity of the human-AI-human collaboration model for future large-scale text mining and knowledge network construction.