AI‑powered Translator & Proofreader app, leveraging multiple context sources for enhanced accuracy, with:
- Translation & proofreading modes
- Tracked changes ingestion (DOCX revisions + TSV)
- Figure / image contextualization (multimodal)
- Exact-match Translation Memory (TMX / TXT)
- Automatic TMX export (Translate mode)
- Chunked processing for large corpora
- Multiple LLMs supported (Claude / Gemini / OpenAI)
- GUI-based operation (Tkinter)
Current version: 2.1.1
| Capability | Translate Mode | Proofread Mode |
|---|---|---|
| Source ingestion (.txt) | 1 column (source) | source{TAB}target{TAB}comment |
| TM (exact match) | Applied pre‑LLM | Not applied |
| Tracked changes context | Yes (relevant subset) | Yes (relevant subset) |
| Images (fig refs) | Injected before referenced lines | Same |
| Output TXT | source{TAB}translated | source{TAB}revised_target{TAB}comment |
| Output TMX | Yes | No |
| Custom instructions | Appended to prompt | Appended to prompt |
| Comments merging | N/A | Original + AI summary (conditional) |
- Python: 3.10+ recommended.
- Create / activate virtual environment (optional):
python -m venv venv source venv/Scripts/activate (Windows: venv\Scripts\activate) - Install dependencies (install only what you need):
pip install google-generativeai anthropic openai pillow - Run:
python Supervertaler_v2.0.0.py
Pillow is optional (only needed for image / multimodal figure context).
If a provider library or API key is missing, related functionality is skipped gracefully.
Create api_keys.txt next to the script (auto-template generated if absent):
#google = YOUR_GOOGLE_API_KEY_HERE
#claude = YOUR_CLAUDE_API_KEY_HERE
#openai = YOUR_OPENAI_API_KEY_HERE
Uncomment and fill lines. Multiple providers can coexist.
Source sentence 1
Source sentence 2
...
If you re‑feed a previously exported 2‑column file:
Source sentence 1{TAB}Old translation
Only the first column is read; the rest are ignored.
Source{TAB}ExistingTarget
Source{TAB}ExistingTarget{TAB}Optional prior comment
Source{TAB}NewTranslation
Plus: *.tmx (exact matches + generated translations, excludes error placeholders).
Source{TAB}RevisedTarget{TAB}MergedComment
MergedComment includes (if present):
- ORIGINAL COMMENT:
- PROOFREADER COMMENT (AI):
If no changes: AI may omit revisions or summarize “No changes”.
Load:
- DOCX with Word tracked revisions (insertions/deletions)
- TSV:
Original{TAB}Final
Browser UI:
- Search (substring or exact)
- View full original / final pair
- Copy single or combined entries
Context Injection:
- Per batch: exact matches first, then partial word-overlap (≥2 meaningful tokens)
- Capped length to protect token budget
Place figure image files in a chosen “Document Images Folder”.
Normalization examples (all map to fig1a):
Figure 1A.png
fig_1-a.JPG
Figuur 1-A.webp
On a line referencing “Figure 1A” / “Fig. 1A” the corresponding image is inserted into the model prompt before that line.
Supported providers:
- Gemini: PIL images embedded in prompt.
- Claude: base64 image blocks.
- OpenAI: data URL image_url blocks.
Supported:
- TMX 1.4 (
<tu><tuv xml:lang=".."><seg>...) - Plain TXT:
source{TAB}target
Process:
- Normalize GUI language codes (e.g., “English”→“en”).
- Apply exact matches first (fills target, skips LLM call).
- Remaining segments go to model.
No fuzzy matching yet (planned).
Free‑text field appended verbatim near the system directive.
Use for:
- Terminology constraints
- Style (e.g., “Use formal patent phrasing”)
- Locale notes
- Chunk Size = number of lines per model request.
- TM filtering reduces model cost where matches exist.
- Each batch includes: needed source lines, optional images, tracked-change subset, and instructions.
- Real‑time log pane (queue-driven).
- Graceful degradation when:
- Missing provider lib
- Missing API key
- Pillow absent (image context disabled)
- Placeholder inserted if model omits a numbered output line.
Planned (Unreleased):
- Fuzzy TM matches
- Glossary enforcement
- JSON run metadata
- Token-based tracked-change scoring
- Enhanced provider capability introspection
- Fork / feature branch.
- Add or update tests (if you introduce logic units).
- Update CHANGELOG.md (Unreleased section).
- Submit PR with concise summary.
(Choose or add a LICENSE file: MIT / Apache-2.0 / Proprietary – not specified yet.)
| Task | Done |
|---|---|
| Install dependencies | ☐ |
| Add API keys | ☐ |
| Prepare input TXT | ☐ |
| (Optional) TMX/TXT TM | ☐ |
| (Optional) DOCX / TSV tracked changes | ☐ |
| (Optional) Images folder | ☐ |
| Run script & select mode | ☐ |
| Review log & outputs | ☐ |
Open an issue with:
- Version (shown at startup)
- Mode (Translate / Proofread)
- Provider + model
- Minimal repro input snippet
- Stack trace (if any)
Happy translating & proofreading!
- Add a PNG named “Figure 1A.png” to your images folder.
- Include “Figure 1A” in a test segment.
- Run with Claude and OpenAI (e.g., claude-3-5-sonnet-20241022, gpt-4o).
- Check Log for “Added Image for Figure Ref …” messages.