Cross-Domain Multilingual Model Using Representation Alignment

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing chat bots require significant time and resources to build large datasets in multiple languages for effective multilingual communication, limiting their ability to provide responses in languages other than English.

Innovation Solution

A cross-domain multilingual model is generated using a system that trains a first model in a single language and a second model in multiple languages, leveraging isotropic regularizers and knowledge distillation to align representations across languages, allowing for efficient deployment in various domains with minimal data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If traditional models train chat bots using large datasets in multiple languages, then multilingual communication capability is improved, but time and resources required for building datasets increase significantly

Engineering Contradiction:

Improvemultilingual communication capabilityVSAvoidtime required for building datasets

Core Design Contradiction:

Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent segments the multilingual training process into two stages: first training a monolingual model on high-quality data for one language, then using that model to generate synthetic data for multiple other languages. This segmentation allows the system to achieve multilingual capability without requiring large datasets for each language simultaneously, thus reducing the time and resources needed for data collection.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent uses the trained monolingual model to generate synthetic data that copies the linguistic patterns and structures of the source language, which are then used to train multilingual models. This copying approach enables the system to create training data for multiple languages from a single language's data, significantly reducing the time required for building multilingual datasets.

Inventive Principle:

Principle #26Copying

2Adaptability or versatility

If traditional models train chat bots using large datasets in multiple languages, then multilingual communication capability is improved, but computational resources and time required for training increase significantly

Engineering Contradiction:

Improvemultilingual communication capabilityVSAvoidcomputational resources required for training

Core Design Contradiction:

Adaptability or versatilityVSUse of energy by moving object

Solution Approach 1:

The patent performs preliminary training of a monolingual model before generating synthetic data for multilingual training. This preliminary action creates a foundation model that can efficiently generate synthetic data for multiple languages, reducing the computational resources needed compared to training separate models for each language from scratch.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system uses the pre-trained monolingual model to copy and generate synthetic training data for multiple languages. This copying approach is computationally more efficient than collecting and processing actual multilingual data, as it leverages the existing monolingual model's parameters to create synthetic data representations, thereby reducing overall computational resource requirements.

Inventive Principle:

Principle #26Copying

3Adaptability or versatility

If chat bots are trained to provide responses in multiple languages, then customer service accessibility is improved, but the complexity of building and maintaining multilingual datasets increases

Engineering Contradiction:

Improveresponse language capabilityVSAvoidcomplexity of building and maintaining datasets

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent implements a self-service approach where the trained monolingual model automatically generates synthetic training data for multiple languages without requiring manual data collection or annotation. This self-service capability simplifies the process of building multilingual datasets, as the system uses itself to create the training data rather than requiring external human resources for data preparation.

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The system copies the linguistic structures and patterns from the source language data to generate synthetic data for target languages. This copying mechanism simplifies dataset maintenance by automatically generating consistent training data that mirrors the quality and structure of the original monolingual data, eliminating the need for manual curation and maintenance of multilingual datasets.

Inventive Principle:

Principle #26Copying

Data Source

PatentUS20260044785A1System and method for generating a cross-domain multilingual model

Publication Date: 2026.02.12 WALMART APOLLO LLC

AI summary

System and methods for generating a cross-domain multilingual model are disclosed. In some embodiments, a disclosed method includes: storing, in a database, a plurality of first utterances associated with a first language, training a first model using the plurality of first utterances, the first model being associated with the first language, generating, using the first model, a plurality of first representations associated with the plurality of first utterances, training a second model, using the plurality of first representations, the second model being associated with a plurality of second languages, receiving, using the second model, a second utterance in the second language, and generating, using the second model, a response in one or more languages of the plurality of second languages.