Alternative CPU and GPU Parallel Computing Approaches for Improving Sequential Analysis of Probability Associations in Short Texts

Abstract

In linguistics, probabilistic relation between co-occurrent words can provide useful interpretation of knowledge conveyed in a text. Connectivity patterns of vectorized representation of lexemes can be identified by using bigram models of word sequences. Similarity assessment of these patterns is performed by applying cosine similarity and mean squared error measures on word vectors of probabilistic relation matrix of text. Moreover, parallel computing is another important aspect for various domains that enables fast data processing and analytics. In this paper, we aim to demonstrate the benefit of parallel computing for computational challenges of extracting probabilistic relations between lexemes. In this study, we have explored performance limitations of sequential semantic similarity analysis and then implemented CPU and GPU parallel versions to show benefits of multicore CPU-GPU utilization for computationally demanding applications. Our results indicate that the alternative parallel computing implementations can be used to significantly enhance performance and applicability of probabilistic relation graph models in linguistic analyses.