以盛唐時期絕句與律詩色彩意象評估大型語言模型文本分類能力之研究-以ChatGPT為例

Abstract

隨著科技進步與人工智慧蓬勃發展，生成式 AI 技術日益受到關注，其與文學文化的結合成為重要研究議題。在深度學習的推動下，基於 Transformer 架構的語言模型如 BERT、ChatGPT、LLaMA 2 等，已成為自然語言處理（NLP）任務的主流工具，並廣泛應用於各專業領域；然而，在文學領域詩詞方面，這些模型的應用仍不多，本研究將探討 ChatGPT 在唐詩色彩意象分類上的可行性與應用潛力，採用小林重順提出的 174 種色彩意象為分類標準，建立一套唐詩色彩意象，為設計應用領域提供參考，並藉此評估 ChatGPT 在文學文本分類任務中的表現。方法上，本研究將 ChatGPT 與專家進行色彩意象分類的結果分別輸出並設計成問卷，由社會大眾進行選擇與評估，藉此比較ChatGPT與專業分類的表現。結果顯示，雖整體而言專家分類更符合大眾期待，但 ChatGPT 仍有約四成唐詩的分類結果獲得較高認同，顯示ChatGPT在部分唐詩的色彩意象判別上具備一定的可行性與準確性；研究進一步彙整問卷回饋，歸納出 14 種盛唐絕句與律詩的色彩意象，包括懷鄉、快樂、寧靜、苦味、自然、夢幻等，作為未來跨域應用的參考依據；最終，這些色彩意象不僅有助於理解唐詩的情感內涵，也反映了盛唐社會的文化氛圍與歷史背景，進而為文案設計、廣告、視覺藝術、空間設計等跨領域創作提供色彩應用上的參考依據，拓展其應用價值與創新可能。

With the rapid advancement of technology and artificial intelligence, generative AI has increasingly drawn attention, especially in its integration with literature and culture, which has become a significant topic of academic research. Driven by deep learning, language models based on the Transformer architecture such as BERT, ChatGPT, and LLaMA 2 have become mainstream tools for natural language processing (NLP) tasks and are widely applied across various professional fields. However, their application in the literary domain, particularly in classical poetry, remains limited. This study explores the feasibility and potential of using ChatGPT to classify color image in Tang poetry, employing the 174 types of color image proposed by Shigekatsu Kobayashi as the classification framework. A dedicated system for color image in Tang poetry is developed to provide references for design-related fields and to assess ChatGPT performance in literary text classification tasks. Methodologically, the study compares color image classifications generated by ChatGPT and those by human experts, presenting both results in a questionnaire for public evaluation. The comparison aims to determine how ChatGPT performance aligns with expert judgment. The results show that, while expert classifications generally align more closely with public expectations, approximately 40% of ChatGPT classifications of Tang poems received higher public approval, indicating a certain degree of feasibility and accuracy in ChatGPT interpretation of poetic color image. Based on survey feedback, the study further identified 14 representative color image categories in golden era of Tang Dynasty quatrains and regulated verses, including themes such as nostalgic, lighthearted, friendly, happy, peaceful, placid, provincial, diligent, bitter, heavy and deep, complex, natural, sweet and dreamy, and amiable. These categories serve as references for future interdisciplinary applications. Ultimately, the identified color image not only aids in interpreting the emotional connotations of Tang poetry but also reflects the cultural and historical context of the golden era of Tang Dynasty, offering valuable reference points for creative work in fields such as copywriting, advertising, visual arts, and spatial design, thus expanding the application value and innovative potential of color image in literary and artistic domains.