g-binary: A New Non-parameterized Code for Improved Inverted File Compression

Abstract

The inverted file is a popular and efficient method for indexing text databases and is being used widely in information retrieval applications. As a result, the research literature is rich in models (global and local) that describe and compress inverted file indexes. Global models compress the entire inverted file index using the same method and can be distinguished in parameterized and non-parameterized ones. The latter utilize fixed codes and are applicable to dynamic collections of documents. Local models are always parameterized in the sense that the method they use makes assumptions about the distribution of each and every word in the document collection of the text database. In the present study, we examine some of the most significant integer compression codes and propose g-binary, a new non-parameterized coding scheme that combines the Golomb codes and the binary representation of integers. The proposed new coding scheme does not introduce any extra computational overhead when compared to the existing non-parameterized codes. With regard to storage utilization efficiency, experimental runs conducted on a number of TREC text database collections reveal an improvement of about 6% over the existing non-parameterized codes. This is an improvement that can make a difference for very large text database collections.