Context Vector Encoding for Mixed-Language Automatic Interpretation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automatic interpretation technologies face challenges in effectively using recognition results across multiple languages due to the nature of automatic interpretation, particularly in situations involving Korean-English interactions, where previous utterances in Korean cannot be directly input with English voice.
Innovation Solution
A multilingual embedding language model is trained to incorporate context vectors from previous utterances, allowing for improved interpretation performance by using a multilingual embedding language model that transforms dimensional spaces to align interpretation target languages and utilizes techniques like masked language models and next sentence prediction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If speech recognition results from previous utterances are to be used for improving interpretation accuracy, then interpretation performance is improved, but it becomes difficult to explicitly use recognition result text because there are two or more different languages
Solution Approach 1:
The patent introduces an intermediary mechanism (context vector encoding and multilingual embedding) that mediates between speech recognition results in different languages. Instead of directly inputting text in one language to interpret another language, the system encodes previous utterances into language-independent context vectors that can be universally applied across language pairs, thus resolving the language barrier while maintaining context utilization
Solution Approach 2:
The patent transforms the parameter representation of previous utterances from language-specific text to language-independent contextual embeddings. By changing the parameter space from textual language data to numerical embedding vectors, the system enables cross-lingual context utilization without being constrained by language differences
2Measurement precision
If a multilingual embedding language model is trained to incorporate context vectors from previous utterances, then interpretation performance in mixed-language scenarios is improved, but device complexity increases
Solution Approach 1:
The patent applies preliminary action by pre-training embedding models on multilingual corpora before the actual interpretation task. This preliminary training phase creates reusable contextual embeddings that can be directly applied during interpretation without requiring complex real-time multilingual processing, thus reducing operational complexity while maintaining high accuracy
Solution Approach 2:
The patent creates a universal multilingual embedding model that serves multiple language pairs and interpretation scenarios simultaneously. This single model handles Korean-English, English-Korean, and other language combinations, eliminating the need for separate models for each language pair and reducing overall system complexity
Data Source
AI summary
An apparatus for improving context-based automatic interpretation performance includes: an uttered voice input unit configured to receive a voice signal from a user; a previous sentence input unit configured to determine whether there is a user's previous utterance when the voice signal is input by the uttered voice input unit; a voice encoding processing unit configured to decode only the voice signal through the uttered voice input unit when it is determined that there is no user's previous utterance and extract a vector of the voice signal when it is determined that there is the user's previous utterance; a context encoding processing unit configured to extract a context vector from a previous utterance when there is the previous utterance and transmit the extracted context vector of the previous utterance; and an interpretation decoding processing unit configured to output an interpretation result text.