The semantic gap is the lack of coincidence between the information one can extract from data and its interpretation. It is an yet solved issue for multimedia applications like image captioning, where it is often challenging to select the best fitting wording out of a group of candidates. To create a measurement for the perceived differences between words, this doctoral research proposes the idea of analyzing crawled image data to gain a better semantic understanding of language and vision. Abstract words have a broad mental image due to them being less visually defined, while concrete words with a rather narrow visual feature space are visually easier to grasp. The core goal of this research is to approximate this perceived abstractness of those words as a metric. The thesis proposes two methods, looking at both relative and absolute measurements. For the relative method, a data-driven approach is proposed, while the absolute measurements train a machine-learned model on existing data from Psycholinguistics. Each method is evaluated using crowd-sourced data and compared to related approaches. With this, the thesis presents methods to analyze the mental image of words from different angles, targeting a way to quantify the semantic gap between vision and language.
Type: Technical Report at CVIM (情報処理学会研究報告) Special Session on Doctoral Theses
Publication date: May 2020