Sentence Generation Method Using Word Set Probability Adjustment

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current machine translation methods face challenges in accurately translating sentences across languages, leading to variations in understanding due to ambiguity and lack of context, particularly when dealing with words that have multiple meanings or require synonyms and basic forms.

Innovation Solution

A sentence generating method that involves generating feature vectors from source sentences, adjusting probabilities of generated words based on source and corresponding word sets, and selecting character strings using a combination of adjusted and unadjusted probabilities to improve translation accuracy by incorporating context and synonyms.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If machine translation is performed using conventional methods, then translation speed is maintained, but translation accuracy deteriorates due to ambiguity and lack of context

Engineering Contradiction:
Improvetranslation accuracyVSAvoidcontext information
Core Design Contradiction:
Measurement precisionVSLoss of information

Solution Approach 1:

The patent applies preliminary action by generating a corresponding word set from the source sentence before performing the actual translation. This pre-processing step creates a reference framework that guides the translation process, ensuring contextually appropriate word selections are made beforehand, thereby improving translation accuracy while preserving context information.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent implements feedback by adjusting the probability of selected words based on whether they appear in the pre-generated corresponding word set. Words that match the corresponding word set receive probability adjustments that favor their selection, creating a feedback loop that continuously refines translation accuracy by referencing contextual information from the source sentence.

Inventive Principle:
Principle #23Feedback

2Measurement precision

If probability adjustment based on word sets is applied, then translation accuracy is improved, but computational complexity increases

Engineering Contradiction:
Improvetranslation accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies segmentation by dividing the translation process into distinct stages: generating the corresponding word set, performing initial translation, and then adjusting probabilities based on set membership. This segmentation allows each stage to be optimized independently, improving overall accuracy while managing computational complexity through structured processing steps.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent uses parameter changes by modifying the probability values of translated words based on their presence in the corresponding word set. This parameter adjustment mechanism dynamically tweaks translation outcomes without requiring complete re-computation, thereby improving accuracy while keeping computational complexity manageable through localized parameter modifications.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentUS10990765B2Sentence generating method and apparatus
Publication Date: 2021.04.27 SAMSUNG ELECTRONICS CO LTD
  • US10990765B2 patent drawing
  • US10990765B2 patent drawing
  • US10990765B2 patent drawing

AI summary

A sentence generating method includes: generating a corresponding word set of a source word set generated based on a source sentence; generating words by performing decoding based on feature vectors generated through encoding of the source sentence; adjusting a probability of at least one of the generated words based either one or both of the source word set and the corresponding word set; and selecting character strings from different character strings including each of the generated words based on the adjusted probability and the probability as unadjusted.