Word graph rescoring method and system for deep learning language models

Abstract

The invention discloses a word graph rescoring method and system for deep learning language models for intelligent conversation voice platforms and electronic equipment. The method comprises the following steps of: storing an output word on each edge or each node of a word graph; traversing each node and each edge of the word graph in sequence from a starting node of the word graph, recording ki paths from the starting node to the ith node, and connecting the output words on all the edges or all the nodes on each path in series to form ki word sequences; carrying out reduction processing on the ki word sequences t form ji word sequences; and calculating scores of ji word sequences of each ith node by calling a deep learning language model. According to the method, the word graph is used asa rescoring target, so that the problem of small search space is solved; the problem of redundant repeated calculation is solved by using history cache; history clustering, token pruning and clusterpruning are used for decreasing extension of the word graph, accelerating the calculation and decreasing the memory consumption; and the word graph rescoring efficiency is improved by adoption of a node parallel calculation.

Description

technical field [0001] The invention belongs to the technical field of language model re-scoring, and in particular relates to a word graph re-scoring method, system and electronic equipment for a deep learning language model of an intelligent dialogue voice platform. Background technique [0002] The re-scoring technology refers to dividing speech recognition and decoding into multiple stages, and then using a language model with better performance to re-evaluate the candidate intermediate results for the results of the first-pass recognition and decoding, and finally obtain a new decoding result. According to the different intermediate results, it can be divided into multi-candidate list re-scoring (N-best Rescoring) and word graph re-scoring (LatticeRescoring). Among them, multi-candidate list re-scoring uses the language model with better performance to re-estimate the language score of the top N candidate results obtained by decoding the first pass of speech recognition...

AI Technical Summary

Problems solved by technology

[0004] In the process of implementing the present invention, the inventors found that due to the structure of multi-candidate list re-scoring, there will be a large amount of redundant information leading to repeated calculations, and the search space is small and the scalability is not strong

Due to its special structure, word graph re-scoring also leads to a huge search space, exponential growth of word graph expansion, large computing resources and memory consumption, and slow speed

The modeling ability of the N-Gram language model is insufficient, resulting in limited improvement in speech recognition performance

Images