隨著高齡化社會發展，遺忘型輕度認知障礙（Amnestic Mild Cognitive Impairment， aMCI）因具有較高機率進展為阿茲海默症，早期偵測與評估逐漸受到重視。過去研究多聚焦於詞彙、句法等微觀語言特徵，較少探討語篇層級之語意結構與主題維持能力，且中文相關研究仍相對有限。 本研究以 DementiaBank 之 Chou 中文語音資料集作為次級資料，分析正常高齡者與遺忘型輕度認知障礙受訪者於三項圖片描述任務中的語篇表現，本研究建立宏觀命題、微觀命題、無關命題及不完整命題等語意結構指標，並進一步分析主題覆蓋率、主題切換次數、離題切換次數及主題重複等主題偏移指標，以探討記憶能力、語意結構與主題偏移之間的關係。除此之外，整合記憶量表分數、語言特徵、語意結構指標及主題偏移指標，建立機器學習模型進行分類評估。 研究結果顯示，不完整命題較能反映認知功能下降之語篇規劃與語意整合可能存在 困難，且微觀命題較高者具有較佳之主題覆蓋能力，而無關命題較多者則較容易出現主題切換與離題現象，顯示語意結構與主題維持能力之間具有密切關聯。機器學習結果亦顯示，整合量表、語言特徵、語意結構及主題偏移指標後，雖提升效果有限，但仍具備良好分類表現，其中Logistic Regression 模型達最佳效能。 本研究從語意結構與主題偏移角度探討中文遺忘型輕度認知障礙者之語篇表現特 徵，除補充中文相關研究之不足外，亦可作為未來臨床評估與早期篩檢之參考依據。;Amnestic Mild Cognitive Impairment (aMCI) is a high-risk stage for the development of Alzheimer′s disease， making early detection increasingly important. Previous studies have mainly focused on lexical and syntactic， while discourse-level semantic organization and topic maintenance have received less attention， particularly in Mandarin Chinese. This study analyzed discourse data from cognitively healthy older adults and individuals with aMCI using the Chou Mandarin corpus from DementiaBank. Semantic structure indicators， including macropropositions， micropropositions， irrelevant propositions， and incomplete propositions， were examined together with topic-shift measures， including topic coverage， topic-switch frequency， off-topic switches， and topic repetition. In addition， memory assessment scores， linguistic features， semantic structure indicators， and topic-shift measures were integrated into machine learning models for classification. The results showed that incomplete propositions were more sensitive to discourse planning and semantic-integration difficulties associated with cognitive decline. Participants with higher microproposition counts demonstrated better topic coverage， whereas those producing more irrelevant propositions exhibited more frequent topic shifts and off-topic transitions. Furthermore， integrating scale-based measures， linguistic features， semantic structural features， and topic-shift indicators resulted in only limited performance gains， but still achieved strong classification performance， with the Logistic Regression demonstrating the best overall performance. This study provides a discourse-based framework for analyzing language performance in Mandarin-speaking individuals with aMCI and offers potential linguistic markers for future clinical assessment and early screening of cognitive impairment.
