Vector Space Representations in Information Retrieval

This publication doesn't include Faculty of Arts. It includes Faculty of Informatics. Official publication website can be found on muni.cz.

Authors

NOVOTNÝ Vít

MU Faculty or unit

Faculty of Informatics

Citation
Description Modern text retrieval systems employ text segmentation during the indexing of documents. I show that, rather than returning the segments to the user, significant improvements are achieved on the semantic text similarity task by combining all segments from a single document into one result with an aggregate similarity score. Standard text retrieval methods underestimate the semantic similarity between documents that use synonymous terms. Latent semantic indexing tackles the problem by clustering frequently co-occuring terms at the cost of the periodical reindexing of dynamic document collections and the suboptimality of co-occurences as a measure of synonymy. I develop a term similarity model that suffers neither of these flaws.
Related projects: