Neural machine translation decoding acceleration method based on discrete variables

Abstract

The invention discloses a neural machine translation decoding acceleration method based on discrete variables. The neural machine translation decoding acceleration method comprises the steps of: constructing a neural machine translation model which is used for training parallel corpora and based on an attention mechanism, and taking a model parameter after training convergence as a baseline system; acquiring a scaling factor (scale) of each tensor in the baseline system through adopting a quantization method, and initializing a unified scaling factor (base_scale) for the whole model; carryingout operation on a neural machine translation model; calculating to obtain a common scaling factor for the scaling factors from different operations before carrying out an addition operation, so as toensure that the magnitudes of the parameters are consistent; and performing inverse quantization on the output of the neural machine translation model, sending a model output result to a normalization function, and acquiring a final translation result. By utilizing the natural advantage of low operation complexity of a fixed point number, the real-time corresponding speed is improved on the premise that the model performance is not degraded on the basis of the latest implementation of rapid reasoning.

Description

technical field [0001] The invention relates to a neural machine translation decoding acceleration technology, in particular to a neural machine translation decoding acceleration method based on discrete variables. Background technique [0002] Machine translation (Machine Translation) is the use of computer programs to translate a natural language into another natural language, which belongs to the category of computational linguistics. In 1949, Warren Weaver published a memorandum titled "Translation", which marked that machine translation based on modern computers officially entered the stage of history. Machine translation not only involves human cognition of its own language and way of thinking, but also involves many fields such as artificial intelligence, information theory, knowledge engineering, and software engineering. It is a discipline that intersects multiple technologies in depth. In the past ten years, the research and industrialization of machine translatio...

AI Technical Summary

Problems solved by technology

[0009] In view of the fact that the machine translation method in the prior art relies too much on single-precision and double-precision floating point and cannot take advantage of low-precision numerical operations, the technical problem to be solved by the present invention is to provide a neural machine translation decoding acceleration method based on discrete variables , making full use of the natural advantages of low computational complexity of fixed-point numbers, based on the latest implementation of fast reasoning, and on the premise of almost no decline in model performance, the real-time response speed can be improved

Images