Multilingual Search Query Translation for ML Model Precision
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing information search systems experience differences in search performance when querying in the same language versus translated languages, leading to discrepancies in search results, particularly when searching Japanese documents with English queries.
Innovation Solution
An information searching program and apparatus that translates search queries to and from a primary language, using a machine learning model trained in that language, to standardize the search process and reduce language-related discrepancies, by translating search conditions to the primary language for processing and re-translating results as needed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If search queries are directly processed in their original language without translation, then processing speed is maintained, but search precision deteriorates due to language-related discrepancies in the machine learning model
Solution Approach 1:
The patent applies preliminary action by translating search queries to the training language (Japanese) before processing them through the machine learning model. This pre-translation step ensures that all queries, regardless of original language, are converted to the language the model was trained on, thereby improving search precision without requiring changes to the model itself.
Solution Approach 2:
The patent uses translation as an intermediary process between the user's query language and the machine learning model's processing language. The translation component acts as a mediator that converts queries from various languages into Japanese, enabling the model to process them uniformly and improving overall search accuracy across multilingual inputs.
2Adaptability or versatility
If the machine learning model is trained on multiple languages, then adaptability to different languages improves, but manufacturing precision deteriorates due to increased complexity in training data preparation
Solution Approach 1:
The patent achieves universality by designing a translation-based preprocessing system that handles multiple languages through a single unified approach. Instead of training separate models for each language or preparing separate training datasets for each language, the system translates all queries to Japanese and uses a single Japanese-trained model, thereby supporting multiple languages while maintaining training data precision.
Solution Approach 2:
The translation component serves as an intermediary that enables the Japanese-trained model to handle queries in various languages. This approach allows the system to maintain high training data precision by using only Japanese training data while still achieving broad language adaptability through the translation mediator.
3Measurement precision
If translation is performed for all search queries, then language-related discrepancies are reduced, but device complexity increases due to additional translation components
Solution Approach 1:
The patent introduces a translation intermediary that converts queries to Japanese before processing. This single translation component resolves language-related discrepancies and improves search result accuracy across all languages, while the rest of the system architecture remains relatively simple and unchanged.
Solution Approach 2:
The patent changes the language parameter of the query from its original language to Japanese through translation. This parameter transformation enables the existing Japanese-trained model to process multilingual queries effectively, improving accuracy without requiring fundamental changes to the system architecture or adding complex multilingual model components.
Data Source
AI summary
A non-transitory computer-readable recording medium stores an information searching program causing a computer to execute processing of: in an information search using a trained machine learning model generated by machine learning using training data in a first language, when a search condition is in the first language, translating the search condition to a second language different from the first language; re-translating the search condition translated to the second language to the first language; and performing the information search by inputting the search condition re-translated to the first language to the machine learning model; and when the search condition is in a language different from the first language, translating the search condition to the first language, and performing the information search by inputting the search condition translated to the first language to the machine learning model.