Distributed Model Partitioning for Scalable Machine Translation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current machine translation systems face challenges in achieving high translation quality and speed due to limitations in handling large language models and varying workloads, leading to suboptimal performance in scalability and adaptability.

Innovation Solution

The implementation of distributed machine processing systems using resource partitioning, replication, and load balancing techniques allows for scalable and adaptive processing, enabling efficient access to large models across multiple servers, thereby improving translation quality and speed while handling high volumes of requests.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If a single server stores and processes the entire language model, then the system structure is simple, but the system cannot handle large volumes of requests and has poor scalability

Engineering Contradiction:
ImprovescalabilityVSAvoidsystem structure
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The language model is divided into multiple partitions, with each partition stored on a separate data server. This segmentation allows the system to distribute the computational load across multiple servers, improving scalability and the ability to handle large volumes of requests while maintaining a manageable system structure through modular organization.

Inventive Principle:
Principle #1Segmentation

2Productivity

If the language model is distributed across multiple servers, then the system can handle large volumes of requests, but accessing the model becomes more complex

Engineering Contradiction:
Improverequest handling capacityVSAvoiddata access complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

A processing server acts as an intermediary between the client and the distributed data servers. The processing server receives translation requests, determines which language model partitions are needed, retrieves the relevant partitions from the appropriate data servers, and assembles the complete language model for processing. This intermediary approach enables the system to handle large volumes of requests by distributing data access while managing the complexity of accessing distributed model partitions through a unified interface.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Speed

If the entire language model is loaded into memory, then access speed is fast, but the memory requirement becomes prohibitively large

Engineering Contradiction:
Improvemodel access speedVSAvoidmemory capacity
Core Design Contradiction:
SpeedVSQuantity of substance

Solution Approach 1:

The language model is segmented into multiple partitions that are distributed across different data servers and stored on disk rather than loaded entirely into memory. When a translation request arrives, the processing server loads only the specific partitions needed for that request into memory, thereby maintaining fast access speed for the required data while significantly reducing the total memory capacity required compared to loading the entire model.

Inventive Principle:
Principle #1Segmentation

4Measurement precision

If manual translation is used, then translation quality is high, but the processing speed is slow and cost is high

Engineering Contradiction:
Improvetranslation qualityVSAvoidprocessing speed
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The system replaces manual mechanical translation with an automated statistical machine translation system that uses a distributed language model. The processing server automatically retrieves relevant language model partitions, applies statistical translation algorithms to translate the source text, and produces the target text without human intervention. This substitution maintains high translation quality through statistical methods while dramatically improving processing speed and reducing costs by eliminating manual labor.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Data Source

PatentEP3534277B1Encoding and adaptive, scalable accessing of distributed models
Publication Date: 2021.10.06 GOOGLE LLC
  • EP3534277B1 patent drawingFigure 1
  • EP3534277B1 patent drawingFigure 2
  • EP3534277B1 patent drawingFigure 3A~3B

AI summary

Systems, methods, and apparatus for accessing distributed models in automated machine processing, including using large language models in machine translation, speech recognition and other applications.