Nowadays, biomedical researchers publish thousands of papers and journals every day. Searching through biomedical literature to keep up with the state of the art is a task of increasing difficulty for many individual researchers. The continuously increasing amount of biomedical text data has resulted in high demands for an efficient and effective biomedical information retrieval (BIR) system. Though many existing information retrieval techniques can be directly applied in BIR, BIR distinguishes itself in the extensive use of biomedical terms and abbreviations which present high ambiguity.;First of all, we studied a fundamental yet simpler problem of word semantic similarity. We proposed a novel semantic word similarity algorithm and related tools called Weighted Edge Similarity Tools (WEST). WEST was motivated by our discovery that humans are more sensitive to the semantic difference due to the categorization than that due to the generalization/specification. Unlike most existing methods which model the semantic similarity of words based on either the depth of their Lowest Common Ancestor (LCA) or the traversal distance of between the word pair in WordNet, WEST also considers the joint contribution of the weighted distance between two words and the weighted depth of their LCA in WordNet. Experiments show that weighted edge based word similarity method has achieved 83.5% accuracy to human judgments.;Query expansion problem can be viewed as selecting top k words which have the maximum accumulated similarity to a given word set. It has been proved as an effective method in BIR and has been studied for over two decades. However, most of the previous researches focus on only one controlled vocabulary: MeSH. In addition, early studies find that applying ontology won't necessarily improve searching performance. In this dissertation, we propose a novel graph based query expansion approach which is able to take advantage of the global information from multiple controlled vocabularies via building a biomedical ontology graph from selected vocabularies in Metathesaurus. We apply Personalized PageRank algorithm on the ontology graph to rank and identify top terms which are highly relevant to the original user query, yet not presented in that query. Those new terms are reordered by a weighted scheme to prioritize specialized concepts. We multiply a scaling factor to those final selected terms to prevent query drifting and append them to the original query in the search. Experiments show that our approach achieves 17.7% improvement in 11 points average precision and recall value against Lucene's default indexing and searching strategy and by 24.8% better against all the other strategies on average. Furthermore, we observe that expanding with specialized concepts rather than generalized concepts can substantially improve the recall-precision performance.;Furthermore, we have successfully applied WEST from the underlying WordNet graph to biomedical ontology graph constructed by multiple controlled vocabularies in Metathesaurus. Experiments indicate that WEST further improve the recall-precision performance.;Finally, we have developed a Graph-based Biomedical Search Engine (G-Bean) for retrieving and visualizing information from literature using our proposed query expansion algorithm. G-Bean accepts any medical related user query and processes them with expanded medical query to search for the MEDLINE database.
展开▼
