TY - BOOK

T1 - Automatic understanding of multimodal content for Web-based learning

AU - Otto, Christian Ralf

PY - 2023

Y1 - 2023

N2 - Web-based learning has become an integral part of everyday life for all ages and backgrounds. On the one hand, the advantages of this learning type, such as availability, accessibility, flexibility, and cost, are apparent. On the other hand, the oversupply of content can lead to learners struggling to find optimal resources efficiently. The interdisciplinary research field Search as Learning is concerned with the analysis and improvement of Web-based learning processes, both on the learner and the computer science side. So far, automatic approaches that assess and recommend learning resources in Search as Learning (SAL) focus on textual, resource, and behavioral features. However, these approaches commonly ignore multimodal aspects. This work addresses this research gap by proposing several approaches that address the question of how multimodal retrieval methods can help support learning on the Web. First, we evaluate whether textual metadata of the TIB AV-Portal can be exploited and enriched by semantic word embeddings to generate video recommendations and, in addition, a video summarization technique to improve exploratory search. Then we turn to the challenging task of knowledge gain prediction that estimates the potential learning success given a specific learning resource. We used data from two user studies for our approaches. The first one observes the knowledge gain when learning with videos in a Massive Open Online Course (MOOC) setting, while the second one provides an informal Web-based learning setting where the subjects have unrestricted access to the Internet. We then extend the purely textual features to include visual, audio, and cross-modal features for a holistic representation of learning resources. By correlating these features with the achieved knowledge gain, we can estimate the impact of a particular learning resource on learning success. We further investigate the influence of multimodal data on the learning process by examining how the combination of visual and textual content generally conveys information. For this purpose, we draw on work from linguistics and visual communications, which investigated the relationship between image and text by means of different metrics and categorizations for several decades. We concretize these metrics to enable their compatibility for machine learning purposes. This process includes the derivation of semantic image-text classes from these metrics. We evaluate all proposals with comprehensive experiments and discuss their impacts and limitations at the end of the thesis

AB - Web-based learning has become an integral part of everyday life for all ages and backgrounds. On the one hand, the advantages of this learning type, such as availability, accessibility, flexibility, and cost, are apparent. On the other hand, the oversupply of content can lead to learners struggling to find optimal resources efficiently. The interdisciplinary research field Search as Learning is concerned with the analysis and improvement of Web-based learning processes, both on the learner and the computer science side. So far, automatic approaches that assess and recommend learning resources in Search as Learning (SAL) focus on textual, resource, and behavioral features. However, these approaches commonly ignore multimodal aspects. This work addresses this research gap by proposing several approaches that address the question of how multimodal retrieval methods can help support learning on the Web. First, we evaluate whether textual metadata of the TIB AV-Portal can be exploited and enriched by semantic word embeddings to generate video recommendations and, in addition, a video summarization technique to improve exploratory search. Then we turn to the challenging task of knowledge gain prediction that estimates the potential learning success given a specific learning resource. We used data from two user studies for our approaches. The first one observes the knowledge gain when learning with videos in a Massive Open Online Course (MOOC) setting, while the second one provides an informal Web-based learning setting where the subjects have unrestricted access to the Internet. We then extend the purely textual features to include visual, audio, and cross-modal features for a holistic representation of learning resources. By correlating these features with the achieved knowledge gain, we can estimate the impact of a particular learning resource on learning success. We further investigate the influence of multimodal data on the learning process by examining how the combination of visual and textual content generally conveys information. For this purpose, we draw on work from linguistics and visual communications, which investigated the relationship between image and text by means of different metrics and categorizations for several decades. We concretize these metrics to enable their compatibility for machine learning purposes. This process includes the derivation of semantic image-text classes from these metrics. We evaluate all proposals with comprehensive experiments and discuss their impacts and limitations at the end of the thesis

U2 - 10.15488/13887

DO - 10.15488/13887

M3 - Doctoral thesis

CY - Hannover

ER -