SheepShuttle Bilingual Data Workflow

SheepShuttle provides comprehensive features for handling bilingual files within SheepComb. You can progress step by step through the top navigation menu.

Step	Mode	Description
1a	Data Extraction	Extract bilingual segments from XLIFF, TMX, TBX, Excel, and CSV files.
1b	Filter & Sampling	Remove duplicate lines, locked segments, or DNT phrases, and generate randomized evaluation samples.
2	Analysis & Structuring	Cross-reference translation memories (TM) and termbases (TB) and calculate Weighted Word Counts (WWC).
3	Data Management & QA	Inspect matches, perform automated QA checks (numbers, tags, terminology), and export data.
4	Build	Re-inject adjusted bilingual data back into original XLIFF files.
5	AI Request (API)	Send chunked segments and tailored prompts to AI models for batch translation and verification.

Detailed Step-by-Step Guide

1a: Data Extraction

Start by extracting translation segments from your source files.

1. Select Extract from the menu.
2. Drag and drop your file into the top-left upload area, or click to browse.
3. Click the Run Extraction button.

Line Breaks in Cells

When extracting from Excel or XLIFF containing internal line breaks, segments are split by line break by default. Uncheck "Split by line breaks" if you prefer keeping cells intact.

Supported Formats

• XLIFF Formats: .xliff, .xlf, .sdlxliff, .mxliff, .mqxliff
• TM & Glossaries: .tmx, .tbx
• Spreadsheets / Tables: .xlsx (Col A: Source, Col B: Target, Col C+: Notes), .csv, .tsv, Word tables, JSON/JSONL

Volume Counts & Export

Above the table, approximate character and word counts are displayed. Click Chara or Word to switch metrics. Use the CSV or JSON buttons at the top right to download extracted data.

1b: Data Exclusion & Filtering

After extraction, the Action Card reveals Filter and Sampling options.

Feature	Description
Remove Duplicate Rows	Automatically eliminates identical source-target pairs beyond the first occurrence.
Remove LOCKED Rows	Excludes locked segments (available for .mxliff / .mqxliff).
DNT (Do Not Translate) Filter	Identifies and filters out non-translatable segments based on selected criteria.

Sampling Evaluation

For quality evaluations, extract representative subsets by specifying a target character count and random seed.

2: Analysis & Structuring

Convert extracted data into standardized project data (ShWvData), cross-referencing translation memories (TM) and termbases (TB).

1. Select Analysis / Structuring from the menu.
2. Add translation memory or glossary files if available.
3. Specify project metadata (Project Name, Source Language, Target Language).
4. Click Run Analysis & Structuring.

Weighted Word Count (WWC)

Calculates workload metrics based on fuzzy match bands:

Supported TM & TB Formats

• TM: .tmx, .xlf, .sdlxliff, .mxliff, .mqxliff, .csv, .xlsx, .json, .jsonl
• TB: .tbx, .xlf, .sdlxliff, .mxliff, .mqxliff, .csv, .xlsx, .json, .jsonl

3: Data Management & QA

Visual Inspection

Inspect matching status with TB terms and TM entries directly in the table:

Click any record to inspect detailed match attributes:

Automated QA Verification

Detect inconsistencies and errors across segments:

QA Category	Description
Number Mismatch	Flags discrepancies in half-width Arabic numerals between source and target.
Tag / Placeholder	Verifies matching pairs of tags (<>, {}, []).
Termbase (TB) Consistency	Checks if approved glossary translations are correctly used.
Inconsistency (100% Matches)	Flags identical source segments with differing translations.
Unedited Post-Edit Leaks	Detects unchanged draft translations where sibling segments were modified.

File Conversion & Export

Export project datasets or prepare chunked files for AI processing:

• JSON Export / Import: Save full project state (.json) or resume previous sessions.
• AI Chunks: Split by file boundary or specific character limits.
• JSONL Export / Import: Export for LLM pipelines and re-import completed outputs.

4: Build

Reconstruct translated bilingual files into CAT-ready formats (e.g., XLIFF):

Select the original bilingual XLIFF and the structured project dataset, then click Start Build to export XLIFF files with updated <target> tags.

5: AI (LLM) Integration

Integrate with SheepBobbin to run batch AI translation or linguistic verification.

1. Select API from the menu.
2. Enter the connection password to establish communication with SheepBobbin.

3. Click Create Chunks to divide dataset into manageable blocks.

4. Configure task parameters:
   • Task Type: Check, Translate, Proof, Diff, or Custom.
   • Languages: Replaces {source_lang} and {target_lang} in prompt templates.
   • Prompt: Customize instructions or load saved templates.
5. Click Process All Chunks to send requests sequentially.
6. Review AI findings and suggestions:

7. Click Export CSV to download results.

Re-running Single Chunks

Click the refresh button on individual chunk headers to re-execute single blocks.

Task Types Reference

Type	Payload Contents	Recommended Use Case
Check	src, tgt, note	AI translation review without TM/glossary noise
Translate	src, note	New translations or drafting from reference
Proof	tgt	Monolingual fluency check
Diff	src, tagged tgt, note	Revision evaluation
Custom	Customizable (src, tgt, note, history, terms)	Advanced contextual prompting

TOON Support (Experimental)

While JSONL is a well-suited format for LLMs, it often leads to higher token usage due to repetitive key names. To address this, TOON (Token-Oriented Object Notation) was developed. TOON is a data format designed specifically for LLMs, expected to reduce token usage by approximately 30–60% compared to JSON.

SheepShuttle supports chunk creation and API requests in TOON format.

TOON support is an experimental feature

While confirmed to work reliably for standard translation and review tasks, complex requests with extensive contextual data—such as history or terminology—have not been fully verified. We recommend starting with simpler use cases such as checking and proofreading.