以眼動儀為工具探討圖像相關判斷準則與相關判斷評分

Abstract

搜尋資訊早已是現代人的日常，而搜尋行為中一般都會經歷「相關判斷」的過程，以確認搜尋的結果是否符合期望。當使用者在檢索文本資料時，經常透過書名、作者、關鍵字等文本內容進行檢索；然而在檢索圖像資料時，卻未必有文本內容可供參考，以致搜尋圖像的品質往往不如預期。因此，相較於傳統文字檢索相關判斷的研究發展，圖像檢索相關判斷的研究起步較晚，目前仍有很大的發展空間。

本研究旨在探討圖像相關判斷準則與建構準則模型，共邀請40位來自於不同系所的大學生作為研究對象。為有效進行相關實驗，本研究建立「判斷圖像→眼動追蹤→圖像評分」的人機互動模式，藉以探討受測者觀看圖像(包括AI生成圖像)所使用的相關判斷準則。此外，本研究採用半結構訪談、圖像相關判斷評分問卷、以及眼動追蹤實驗等三種方法，交互驗證研究過程所得之訪談資料、問卷資料、與實驗數據，以確保研究品質。

本研究主要成果包含三大面向。首先，本研究根據文獻中與搜尋資訊相關的判斷準則，以不同於傳統的實驗方法，經歸納分析後提出適合搜尋圖像之四大類判斷準則，包括：(1)判斷者、(2)判斷情境、(3)展現格式、(4)主觀直覺與客觀特性，並據此建立圖像相關判斷準則的模型，可作為未來相關研究之理論基礎。其次，本研究將圖像展現格式準則分為一般圖像與人工智慧生成式圖像，深入探討受測者判斷AI生成圖像的準則，並納入圖像相關判斷準則與模型。隨著人工智慧技術蓬勃發展，本項成果使得圖像相關性判斷的準則更具完整性與前瞻性。最後，本研究嘗試在此一領域引進眼動儀設備，以較客觀的指標測量值與質化的相關判斷準則的結果進行分析，並探討相關判斷準則與眼動指標測量值的關係，可大幅提升研究結果的嚴謹度。

Information searching has become an essential part of people's everyday lives. Typically, this involves the process of "relevance judgment" to ascertain whether search results meet expectations. When users search for text-based information, there are often references such as book titles, authors, and keywords that can aid the process. However, for image-based searches, these references are not always available, which frequently leads to subpar search qualities. As such, compared to research centering on traditional text-based information searching, research on image-based information searching has progressed rather recently, with many territories left uncharted.

This research primarily focuses on criteria for image relevance judgment and on building a corresponding model, based on a study of 40 college students across various majors. To facilitate effective experimentations, this research structured a human-to-machine interaction mode that involves the process of "image judgment → eye tracking → image rating." This allows the examination of the participants' criteria for image judgment (including AI-generated images). Furthermore, this research employs semi-structured interviews, eye-tracking experiments, and rating questionnaires to assess image relevance, and the results and data are cross-referenced to ensure research quality.

The findings are distinguished into three categories. The first is a set of criteria created from a non-conventional approach based on literature reviews regarding information searching, which, upon utilizing inductive analysis, was organized into four criteria: (1) the judge (i.e., the subject), (2) the judgment context, (3) the presentation format, and (4) subjective intuition and objective characteristics. This set of criteria is then used to construct a judgment model for image relevance standards, which could be used as a foundation for future relevant research. The second category involves classifying the type of image displayed as regular or AI-generated. The subjects were then examined to determine how an image was identified as AI-generated, and the results were incorporated into the aforementioned criteria and model. As AI technology progresses exponentially, this research greatly enhances the comprehensiveness and future-oriented nature of image relevance assessments. Lastly, this research explores an alternative approach by attempting to utilize eye-tracking equipment, thereby producing objective measurements and qualitative results for relevance judgment standards in the analysis. The examination of the relationship between relevance judgment standards and eye-tracking indicator measurements substantially strengthens the rigor and scientific soundness of the research findings.