Cross-lingual Language Model Initialization via Machine Translation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Automated Speech Recognition (ASR) engines face inaccuracies when using language models developed for one context to recognize utterances in different contexts, especially for infrequently used languages or contexts where appropriate models are not available.

Innovation Solution

The method involves machine-translating existing corpora associated with a given language and context to generate a target corpus for a different language and context, estimating a new language model using the translated corpus, and blending it with existing data to improve recognition accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If a language model developed for a given language and context is used to recognize utterances in a different context, then the ASR engine can operate with existing models, but the recognition accuracy deteriorates

Engineering Contradiction:
Improveoperational capabilityVSAvoidrecognition accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent changes the language parameter of the corpus by machine-translating it from a source language to a target language, allowing the ASR engine to adapt language models for different languages and contexts without requiring separate manual development for each combination

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent performs preliminary machine-translation of the corpus to generate a target language corpus before building the language model, so that when the ASR engine needs to recognize utterances in the target language, the language model is already prepared and available

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If a language model specific to target language and context is generated through machine-translation, then recognition accuracy improves, but the device complexity increases

Engineering Contradiction:
Improverecognition accuracyVSAvoidmodel development complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent introduces machine-translation as an intermediary process that converts source language corpora into target language corpora, automating what would otherwise require manual translation and model development efforts for each target language

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent creates a universal approach where a single source language corpus can be machine-translated to generate language models for multiple target languages, making the system multi-functional across different languages without requiring separate development for each

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Ease of manufacture

If machine-translation is used to generate target corpus, then the need for extensive manual development is reduced, but the loss of information may increase due to translation inaccuracies

Engineering Contradiction:
Improvemodel development easeVSAvoidtranslation accuracy
Core Design Contradiction:
Ease of manufactureVSLoss of information

Solution Approach 1:

The patent uses machine-translation to create a copied version of the source language corpus in the target language, which serves as a foundation for building the language model, balancing automation benefits with acceptable translation quality

Inventive Principle:
Principle #26Copying

Data Source

PatentEP3355301B1Cross-lingual initialization of language models
Publication Date: 2022.11.02 GOOGLE LLC
  • EP3355301B1 patent drawingFigure 1
  • EP3355301B1 patent drawingFigure 2
  • EP3355301B1 patent drawingFigure 3

AI summary

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for initializing language models for automatic speech recognition. In one aspect, a method includes receiving logged speech recognition results from an existing corpus that is specific to a given language and a target context, generating a target corpus by machine-translating the logged speech recognition results from the given language to a different, target language, and estimating a language model that is specific to the different, target language and the same, target context, using the target corpus.