UBC Theses and Dissertations
Latent semantic analysis for retrieving related biomedical articles Lin, Sheng-Ting
Retrieving relevant scientific papers in a scalable way is increasingly important, as more and more studies are published. PubMed’s relevant article recommendation is based on MeSH assignments by indexers, which requires significant human resources and can become a limitation in making papers searchable. Many recommendation systems use singular value decomposition (SVD) to pre-compute related products. In this study, we look at using latent semantic analysis (LSA), an application of SVD to determine relationships in a set of documents and terms, to find related biomedical papers. We focused on determining the best parameters for SVD in retrieving relevant biomedical articles given a paper of interest. Using PubMed's recommendations as guidance, we found that using cosine distance to measure document similarity leads to better results than using Euclidean distance. We re-evaluated other parameters, including the weighting scheme and the number of singular values and using a larger abstract corpus. Finally, we asked people to compare the relevant abstract retrieved with our method against those retrieved by PubMed. Our method retrieved sensible articles that were chosen over PubMed's relevant papers one-third of the time. We looked into the abstracts retrieved by either method and discuss possible areas for experimentation and improvement.
Item Citations and Data
Attribution-NonCommercial-ShareAlike 4.0 International