Parallelizing Latent Semantic Indexing using an FPGA-based architecture

Abstract

Latent Semantic Indexing (LSI) has played a significant role in discovering patterns on the relationships between query terms and unstructured documents. However, the inherent characteristics of complex matrix factorization in LSI make it difficult to meet stringent performance requirements. In this paper, we present a deeply pipelined reconfigurable architecture for LSI, which parallelizes the matrix factorization and dimensionality reduction, computation of cosine similarity between vectors, and the ranking of documents. Our architecture implements the reduced Singular Value Decomposition with Hestenes-Jacobi algorithm, in which both singular values and orthogonal vectors are collected, and its components can be reconfigured to update query vector coordinate and calculate query-document similarity. In addition, an ordered tree structure is used to reduce the matrix dimension and rank the documents. Analysis of our design indicates the potential to achieve a performance of 8.9 GFLOPS with dimension-dependent speedups over an optimized software implementation that range from 3.8× to 10.1× in terms of computation time.