RPA-Prompt-TranslationQA

Manual translation QA processes can be time-consuming and inconsistent, especially across large-scale multilingual projects.

This project showcases a real-world example of an AI-powered translation quality assurance (QA) flow built using Microsoft Power Automate and Azure OpenAI (GPT).

It is designed to automatically detect and classify translation errors with structured output, based on carefully engineered prompts.

---

🚀 Overview

• Platform: Microsoft Power Automate
• AI: Azure OpenAI (GPT-4o)
• Use Case: Translation QA Automation
• Input: Excel file with source and translated text
• Output: Structured classification and comparison results written back to Excel

🏆 By implementing this tool in a real production workflow, we achieved a reduction of over 70% in manual translation QA effort.

💡 Since there was no prior knowledge of Power Automate within the team, I independently researched the platform and designed the entire workflow from scratch.

---

📊 Workflow Summary

For each row in the Excel file, the tool constructs a prompt using the Source and Target texts and sends it to GPT in two stages.

1. First GPT call: Performs three tasks in one prompt:

   • Classifies error severity (No error, Minor, etc.)
   • Identifies error type (Mistranslation, Grammar error, etc.)
   • Returns a literal back translation of the Target

2. Second GPT call: Compares the Source and Back Translation to evaluate semantic alignment:

   • Labels the result as Same, Slightly different, or Totally different
   • If no error was previously detected but meanings differ, the result is flagged as a Possible error

All results are written into predefined columns of the Excel template.

---

🧠 Prompt Design Highlight

To minimize the number of API calls and reduce operational cost, the flow is designed so that a single GPT call handles three distinct tasks—error severity classification, error type classification, and back translation—in one structured prompt. The results are returned in a parseable format using delimiters, ensuring efficiency without compromising output clarity.

This prompt is not just an instruction—it defines a workflow. It reflects a shift from using GPT as a text generator to leveraging it as a structured evaluator.

In addition, even when no explicit error is detected, the tool compares the source and back-translated text to catch hidden mismatches. If a significant difference is found, it flags the result as a Possible error, providing an extra layer of QA coverage.

The system prompt includes:

• Clear multi-step instructions
• Strict output format using %%% as delimiter
• Definitions for severity and error types
• Context markers using ##COMM_START## and ##COMM_END##

This demonstrates:

• Practical prompt design for structured, parseable responses
• Real-world integration into an RPA (Robotic Process Automation) workflow

---

📝 Prompt Design Example

Below is a screenshot showing how the prompt is structured and embedded in the Power Automate flow.
It defines three evaluation tasks and uses structured delimiters for consistent parsing:

---

📋 Column Descriptions (Output Fields)

The following columns are pre-defined in the Excel template and are populated during the flow execution:

• ID: Unique identifier for each translation pair
• Source: Original English text
• Target: Translated text (to be evaluated)
• Error Severity: Classification of translation error └─ No error, Minor, Major, Critical, or Possible error (added when Source Comparison reveals a hidden mismatch despite no detected error)
• Error Type: Type of translation error └─ No error, Mistranslation, Opposite meaning, Grammar error, Misspelling, Number mismatch, Other type
• Back Translation: GPT-retranslated English version of the Target
• Source Comparison: GPT evaluation of meaning alignment between Source and back-translation └─ Same, Slightly different, Totally different

---

📈 Sample Output Report

Below is a screenshot of the actual Excel output after the tool has processed several translation pairs:

---

📂 Files

File	Description
definition_example.json	Sanitized JSON definition of the Power Automate flow (not executable)
README.md	This documentation

⚠️ Note: All sensitive data (API keys, user info) has been redacted. This file is shared solely for educational and demonstration purposes. It cannot be executed without customization.

---