用於歷史中文詞義消歧的動態式少量示例提示法

Abstract

本研究提出一種利用大型語言模型（LLM）與動態少量示例提示（Dynamic Few-Shot Prompting）進行歷史漢語語義變化分析的新方法。該方法基於上下文學習（In-Context Learning）框架，根據輸入語句動態選取最具相關性的示例，以實現詞義消歧（Word Sense Disambiguation, WSD）。為了適應古文中特有的詞彙與語境，我們整合了 CTEXT、CBETA、PTT 等多種資料來源，建構出一套涵蓋廣泛的歷史漢語語料庫。實驗結果顯示，在特定條件下，動態少量示例提示相較於零示例與固定示例提示方法展現出更穩定且優異的表現。動態提示的優勢在大型模型中尤為顯著，而在小型模型中則以零示例方法表現更佳。進一步將本方法應用於「家」一詞的語義演變分析，可視化結果揭示該詞在歷代中呈現出多種語義並存與競合的變遷現象，且不同時期呈現出語義間的平衡變化。

This study proposes a novel approach to word sense disambiguation (WSD) for analyzing semantic shifts in the Chinese language from antiquity to the modern era, leveraging general-purpose large language models (LLMs). Specifically, we employ decoder-based models (e.g., GPT) within the framework of in-context learning to design a dynamic few-shot prompting method. This method retrieves semantically relevant examples based on the input context and incorporates them into prompts.

Using an evaluation dataset constructed from the MoeDict dictionary, we conduct comparative experiments between a vector similarity-based approach using BERT and our prompt-based approach using GPT models. The results show that dynamic few-shot prompting achieves high accuracy and robustness in large-scale models, while zero-shot prompting performs better in smaller models.

Furthermore, we apply our method to a comprehensive historical Chinese corpus, constructed by integrating multiple sources such as CTEXT, CBETA, and PTT, and analyze diachronic usage examples of the word jia (家). The visualization of its semantic evolution reveals the coexistence and competition of multiple senses across different historical periods.