AI translation solution presented through a real prompter

Executive Summary

This Case Study aims to discuss our in-house developed AI-based real-time translation solution designed to overcome communication difficulties in corporate workshops with a multilingual environment. Although the initial model was built upon the OpenAI API, it faced issues with mistranslating company-specific terminology and employee names. To address this, we integrated our independently developed Context-Aware Knowledge Base, which significantly improved translation accuracy and achieved low latency, enabling seamless two-way communication. This document introduces our successful case.

The Challenge

The number of workshops involving employees from various language backgrounds within the company has increased. However, language barriers hindered idea sharing and caused some participants to feel excluded, posing a significant obstacle to achieving the fundamental goals of the workshops. Specifically, in offline workshops, real-time comprehension of speakers' presentations and immediate two-way feedback proved difficult.

Key Issues

• Lack of Real-time Communication: Consecutive interpretation interrupted the flow of workshops and reduced participant concentration.
• Information Distortion: There was a possibility that subtle nuances or core information could be omitted or distorted during the interpretation process. Particularly, when company-specific information or employee names were incorrectly translated, the credibility of the translation solution was further compromised.
• Imbalanced Participation: Employees not proficient in a specific language had limited comprehension, making it difficult to access information smoothly

The Solution

To address these issues, we built a custom real-time two-way translation system.

Step 1: Initial Model Implementation

First, we integrated OpenAI's powerful language model API with Speech-to-Text (STT) technology. We implemented a system where if a speaker spoke Korean into the microphone, it would instantly display as English text on the screen, and if an English speaker spoke, it would convert to Korean text. This laid the foundation for basic real-time communication.

Step 2: Innovative Improvement - Integration of In-house Knowledge Base (The Breakthrough)

However, the initial model showed limitations in accurately recognizing company-specific information. While it performed translation, it operated without understanding the organization's ‘meaning’ and ‘context’. For instance, it sometimes mistranslated company slogans, internal project names, and employee names with similar pronunciations, leading to awkward situations.

To solve this problem, we developed and applied a unique technology called the 'Context-Aware Knowledge Engine'.

• Knowledge Base Construction: We refined internal data such as employee rosters from the HR system, in-house glossaries, technical documents, and historical slogans to build a knowledge base for AI learning.
• Knowledge Injection: When a translation request occurs, the AI model is designed to not merely perform general translation but to reference the established knowledge base in real-time. This enables the AI to first determine whether a word in context is a company-specific term or a person's name, then produce the most accurate translation result.