Neural Network Text Translation Length Prediction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Translating text using neural networks is inefficient when the length of the translated text is unknown, often requiring significant memory and computing resources.

Innovation Solution

A neural network model that predicts the length of the output translated sentence based on the input sentence, using pre-ordering and post-ordering algorithms to restructure the text according to the target language's word order, and employs autoregressive and non-autoregressive models with transformer architecture to enhance translation accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If the length of translated text is not known, then translation flexibility is maintained, but memory and computing resources are significantly consumed

Engineering Contradiction:
Improvetranslation flexibilityVSAvoidmemory and computing resources
Core Design Contradiction:
Adaptability or versatilityVSQuantity of substance

Solution Approach 1:

The patent applies preliminary action by predicting the length of the translated text before the actual translation process. A neural network model estimates the target sequence length based on the input text, allowing the system to pre-allocate memory and computing resources. This prevents resource exhaustion during translation while maintaining flexibility, as the prediction is made in advance without committing to a fixed translation outcome.

Inventive Principle:
Principle #10Preliminary action

2Quantity of substance

If traditional neural network translation methods are used without length prediction, then translation coverage is maximized, but translation accuracy decreases due to excessive possible outputs

Engineering Contradiction:
Improvetranslation coverageVSAvoidtranslation accuracy
Core Design Contradiction:
Quantity of substanceVSMeasurement precision

Solution Approach 1:

The patent changes the parameter of sequence length from an unknown variable to a predicted value. By using a neural network to estimate the target length, the system transforms the translation problem from generating any-length output to generating output within a specific length range. This parameter change narrows the search space significantly, improving translation accuracy while maintaining adequate coverage through the neural network's probabilistic predictions.

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If word order restructure algorithms are applied, then translation accuracy for target language syntax is improved, but processing time increases

Engineering Contradiction:
Improvetranslation accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent applies preliminary action by determining the target language word order structure before performing the actual translation. The system analyzes the target language's syntactic requirements in advance and prepares the appropriate restructure rules, so that during translation, the model can directly apply these pre-determined structures rather than discovering them during the translation process. This reduces processing time while maintaining high translation accuracy.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentUS20230267285A1Using one or more neural networks to perform text translation
Publication Date: 2023.08.24 NVIDIA CORP
  • US20230267285A1 patent drawing
  • US20230267285A1 patent drawing
  • US20230267285A1 patent drawing

AI summary

Apparatuses, systems, and techniques to translate a text string. In at least one embodiment, a text string is translated by at least, for example, using one or more neural networks to determine a length of a translated text string before a text string is to be translated.