Neural Network Translation of Named Entities via Segmented Models

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing neural machine translation methods face challenges in accurately translating named entities, such as proper nouns and technical terms, due to their lower frequency of appearance in training data, leading to inadequate learning and difficulty in covering all entities.

Innovation Solution

A translation method that uses a neural network-based text translation model and a named entity translation model, where named entities are replaced with placeholders during the translation process, allowing for separate learning and translation of these entities using a specification model and a named entity translation model, respectively.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If a single neural network model is used for general text translation, then the translation process is simple and fast, but the accuracy of named entity translation deteriorates due to low frequency in training data

Engineering Contradiction:

Improvenamed entity translation accuracyVSAvoidtranslation model complexity

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

The translation system is segmented into two specialized models: a text translation model for general text and a named entity translation model for specific entities. This segmentation allows each model to specialize in its domain, improving named entity translation accuracy without requiring the entire system to be overly complex. The models work together through a unified architecture that routes appropriate text segments to the suitable model.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

A specification model acts as an intermediary that identifies and extracts named entities from input text before translation. This intermediary component enables the system to separate general text from named entities, allowing the appropriate translation model to process each type. The specification model serves as a bridge between the input text and the dual-model translation system.

Inventive Principle:

Principle #24Intermediary (Mediator)

2Measurement precision

If named entities are translated using the same model as general text, then the translation system is simple to implement, but translation accuracy for rare entities deteriorates

Engineering Contradiction:

Improvetranslation accuracyVSAvoidsystem implementation complexity

Core Design Contradiction:

Measurement precisionVSEase of manufacture

Solution Approach 1:

The translation system is divided into specialized components: a text translation model for general text and a named entity translation model for specific entities. This segmentation allows each model to specialize in its domain, improving named entity translation accuracy without requiring the entire system to be overly complex. The models work together through a unified architecture that routes appropriate text segments to the suitable model.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The specification model performs preliminary identification and extraction of named entities from the input text before the translation process begins. This preliminary action allows the system to prepare named entities for specialized processing by the named entity translation model, ensuring they receive appropriate attention during translation without adding significant implementation complexity.

Inventive Principle:

Principle #10Preliminary action

3Reliability

If a translation dictionary is prepared in advance for named entities, then translation accuracy for known entities improves, but the system cannot handle unknown entities and the complexity of maintaining the dictionary increases

Engineering Contradiction:

Improvenamed entity translation accuracyVSAvoidcoverage of all named entities

Core Design Contradiction:

ReliabilityVSAdaptability or versatility

Solution Approach 1:

The named entity translation model uses attention mechanisms that dynamically adjust parameters based on the input context and entity characteristics. This allows the model to adapt to different types of named entities and handle unseen entities by learning from contextual patterns rather than relying on a static dictionary. The model can generalize to unknown entities while maintaining high accuracy for known entities.

Inventive Principle:

Principle #35Parameter changes

Solution Approach 2:

The translation system transitions from a static dictionary-based approach to a dynamic neural network model that can adapt to new entities. The named entity translation model continuously learns from training data and can handle previously unseen entities by leveraging contextual information and attention mechanisms, providing both accuracy for known entities and versatility for unknown entities.

Inventive Principle:

Principle #15Dynamics

Data Source

PatentUS11669695B2Translation method, learning method, and non-transitory computer-readable storage medium for storing translation program to translate a named entity based on an attention score using neural network

Publication Date: 2023.06.06 FUJITSU LTD

AI summary

A translation method, implemented by a computer, includes: converting a text written in a first language into a replacement text in which a named entity in the text is replaced with a predetermined character string; translating the replacement text into a second language by using a text translation model which is a neural network; and translating a named entity corresponding to the predetermined character string in the replacement text into the second language by using a named entity translation model which is a neural network.