Exploiting available domain knowledge to improve the retrieval and recommendation of Digital Cultural Heritage materials

Usman, MA

Abstract

Cultural Heritage (CH) institutions, such as museums, have recently embraced computing techniques to digitise CH materials (artefacts, paintings, books etc) and to make accessible those digital representations through their online portals to millions of museum visitors (onsite and remote). This mass availability of digitised materials, however, can lead to information overload. Therefore, ordinary CH online users can find it challenging to access these materials, because they usually have no domain knowledge and also lack the experience of which precise keyword terms to use to search and discover new information.
As an attempt to mitigate the issues explained above, recommender systems and visual search interfaces have been used by millions of users to discover new and relevant to the users’ interests CH materials. A CH recommender system is a system that uses knowledge — content and social — representations assembled from various domain knowledge sources, to generate personalised recommendations of CH materials. Social knowledge representations provide better recommendation quality than content knowledge representations when they have substantial social knowledge such as user-interactions and social tagging in the representation, but they suffer when available information is insufficient (cold start problem and sparsity of social knowledge).
Different approaches have been deployed to address these challenges, for example a hybrid approach that incorporated content directly into a social knowledge representation to provide a recommendation. But this hybrid approach only works well on domains for which specific content knowledge exists which can directly describe an item and is always available and meaningful. The CH domain does not have such rich specific knowledge that can directly describe the content of CH materials, thus limiting the ability to incorporate content directly into the social knowledge representation for CH recommendations.
To address these challenges, this Thesis starts with examining the strengths and weaknesses of content and social knowledge representations in the context of CH recommendations and how these knowledge representations can complement each other to improve the recommendations of CH materials. The identified knowledge gap is bridged through a new hybrid representation approach by integrating the content and social knowledge representations. The effect of knowledge integration is to increase the instances of quality recommendations and improved discovery, and to provide opportunities to users to discover unexpected and liked recommendations of CH materials.
The new integrated and social knowledge representations are used to further develop a dynamic hybrid CH recommender system. The dynamic hybrid system combines the learned integrated knowledge for each CH object with CH object’s social knowledge, and assigns the weights to both integrated and social knowledge representations to control the contributions of each knowledge so that each representation could contribute based on the current user and search status.
A new visual search interface is also described in this thesis, developed as a part of the research work. The search interface enables visual search and exploration across large CH collections by providing an interactive visual summary of the recommended CH items, addressing the challenge of the lack of domain knowledge by online users. User satisfaction evaluation was conducted to measure the user satisfaction level for using a visual search interface for search and exploration of information from the vast collection when compared to the non-visual search interface. The evaluation showed that a user with no domain knowledge prefers using a visual search interface than one with no visual summary presentation, but the result also shows that there is no significant difference for users that have domain knowledge.
The challenges of evaluating CH recommendations are also addressed in the Thesis. The feedback provided by the users, both implicit and explicit, is exploited to measure and reflect the performance of the cultural heritage recommendation methods used. A user study to evaluate both the ground truth measures and integrated knowledge representations is conducted. Throughout the user study, the results obtained show that the hybrid representation produced a better quality recommendation of CH materials when compared with content and social knowledge representations. The social representation does not provide a better high-level recommendation quality compared to the hybrid representation, but it does outperform the hybrid representation in recommending novel CH materials.

Citation

Usman, M. Exploiting available domain knowledge to improve the retrieval and recommendation of Digital Cultural Heritage materials. (Thesis). University of Salford

• Thesis Type: Thesis
• Deposit Date: Oct 5, 2021
• Publicly Available Date: Oct 5, 2021
• Award Date: Aug 1, 2021