Decoding with Finite-State Transducers on GPUs

TL;DR

This paper presents a GPU-based implementation of algorithms for weighted finite automata, significantly accelerating decoding processes in NLP tasks like speech recognition and tagging.

Contribution

The paper introduces a novel GPU implementation of the Viterbi and forward-backward algorithms for weighted finite automata, enabling faster NLP processing.

Findings

Achieved up to 5.2x speedup over serial implementations.

Realized 6093x speedup over OpenFST.

Demonstrated effectiveness across various NLP tasks.

Abstract

Weighted finite automata and transducers (including hidden Markov models and conditional random fields) are widely used in natural language processing (NLP) to perform tasks such as morphological analysis, part-of-speech tagging, chunking, named entity recognition, speech recognition, and others. Parallelizing finite state algorithms on graphics processing units (GPUs) would benefit many areas of NLP. Although researchers have implemented GPU versions of basic graph algorithms, limited previous work, to our knowledge, has been done on GPU algorithms for weighted finite automata. We introduce a GPU implementation of the Viterbi and forward-backward algorithm, achieving decoding speedups of up to 5.2x over our serial implementation running on different computer architectures and 6093x over OpenFST.