Copy the page URI to the clipboard
Khadka, Anita
(2020).
DOI: https://doi.org/10.21954/ou.ro.00011b89
Abstract
With the continuous growth of scientific literature, it is becoming increasingly challenging to discover relevant scientific publications from the plethora of available academic digital libraries. Despite the current scale, important efforts have been achieved towards the research and development of academic search engines, reference management tools, review management platforms, scientometrics systems, and recommender systems that help finding a variety of relevant scientific items, such as publications, books, researchers, grants and events, among others.
This thesis focuses on recommender systems for scientific publications. Existing systems do not always provide the most relevant scientific publications to users, despite they are present in the recommendation space. A common limitation is the lack of access to the full content of the publications when designing the recommendation methods. Solutions are largely based on the exploitation of metadata (e.g., titles, abstracts, lists of references, etc.), but rarely with the text of the publications. Another important limitation is the lack of time awareness. Existing works have not addressed the important scenario of recommending the most recent publications to users, due to the challenge of recommending items for which no ratings (i.e., user preferences) have been yet provided. The lack of evaluation benchmarks also limits the evolution and progress of the field.
This thesis investigates the use of fine-grained forms of citation knowledge, extracted from the full textual content of scientific publications, to enhance recommendations: citation proximity, citation context, citation section, citation graph and citation intention. We design and develop new recommendation methods that incorporate such knowledge, individually and in combination.
By conducting offline evaluations, as well as user studies, we show how the use of citation knowledge does help enhancing the performance of existing recommendation methods when addressing two key tasks: (i) recommending scientific publications for a given work, and (ii) recommending recent scientific publications to a user. Two novel evaluation benchmarks have also been generated and made available for the scientific community.