GSD: Generalized Stochastic Decoding

Ning Gong, Nianmin Yao

Keywords: Natural Language Processing, Decoding Algorithms

Abstract:

Although substantial progress has been made in various text generation tasks, there remains a vast gap between current generations and human languages. One reason is that virtually all decoding methods currently developed are pragmatic to address the text degeneration problem, which exists in both deterministic and stochastic decoding algorithms. So, why text generated from these algorithms are divergent? What is the critical difference between these algorithms? Moreover, is it possible to design a generalized framework where existing decoding algorithms can be naturally connected, uniformly described, and mutually inspired?
In this paper, we try to explore answers to these intriguing questions. Correctly, we propose a generalized decoding framework that can be used to describe and connect existing popular decoding algorithms. Based on the framework, we propose a novel implementation with a distinctive core from existing decoding algorithms. As far as we know, this is the first work trying to propose a generalized framework to bridge these decoding algorithms using formal theorems and concrete implementations. By setting up different conditions, our framework provides infinite space to develop new decoding algorithms. Experiments show that text produced by our method is closest to the characteristics of human languages. Source code and the generated text can be accessed from https://github.com/ginoailab/gsd.git.

One-sentence Summary: A novel work proposing a generalized framework to connect existing decoding algorithms using formal theorems and concrete implementations. By setting up different conditions, our framework provides infinite space to develop new decoding algorithms.