Impact-Based Document Retrieval

Two of the most important aspects contributing to the success of any document retrieval system are the query mechanism and the representation of its auxiliary operational data. The former greatly affects the quality of the retrieval results as well as the speed of the system. The latter reflects the ability of the system to represent its operational data in a compact form that reduces the storage space and enhances query processing speed. Addressing these two aspects, conventional document retrieval systems rely on a range of stored data, such as term and document frequencies, and document lengths. These statistics are often stored in a well-known index data structure, using compression techniques that are bitor byte-oriented. During query processing, the related statistics are decompressed and then involved in some precise and complicated quantitative formula that determines the final retrieval results. This thesis presents a novel approach of impact-based retrieval that has the twin benefits of enhancing both the effectiveness and the efficiency of retrieval. Several variants of impact-based retrieval are described. The most advanced variant does not require any precise similarity formulation as the impacts are defined qualitatively rather than quantitatively. It gives better effectiveness performance relative to several state-of-the-art retrieval techniques. Moreover, it results in the use of only a few distinct impact values, allowing a more compact organisation of the indexes, and faster query processing. To further accelerate query evaluation, several word-oriented index compression schemes have been developed, offering good compression effectiveness and fast decompression. These schemes take the best features of the bitand byte-oriented methods and represent an excellent compromise between them.