Questions about the naturalness of translated text

[anchor554]

I want to ask a question. The text translated automatically still feels very stiff. Can I add prompt words to make the translated content more natural? How can I adjust it? Thanks

[hitchkin]

Yes, you can try to modify the current existing system prompt by manually editing it in modules/translation/base.py (ideally you'd fork the repo beforehand) which - depending on your prompting skills - could worsen, slightly (!) improve or not affect naturalness in a tangible way.

Which model do you normally use (and for which source <-> target lang. pair?)? Some models do perform better for certain lang. pairs.

Otherwise there is not much that can be done there (at least not quickly anyway*), you'll just have to wait for better models to be released.

----*will paste them into a seperate issue for better visibility-----

[anchor554]

I am currently using openai, and the target language is French translated into simplified Chinese. Now the problem is that after I change the prompt word, the text cannot be displayed. I encountered an error and don’t know how to solve it. Please help me

Last login: Tue May 27 01:04:21 on ttys000
anchor@192 ~ % cd Desktop
anchor@192 Desktop % cd fanyi
anchor@192 fanyi % python3 comic.py
qt.qpa.fonts: Populating font family aliases took 115 ms. Replace uses of missing font family "Arial, Sans-serif" with one that exists to avoid this cost.
2025-05-27 01:16:29.462 Python[35291:446267] error messaging the mach port for IMKCFRunLoopWakeUpReliable
2025-05-27 01:16:30.441 Python[35291:446267] The class 'NSOpenPanel' overrides the method identifier. This method is implemented by class 'NSWindow'
2025-05-27 01:17:03.568 | INFO | modules.utils.inpainting:load_jit_model:110 - Loading model from: /Users/anchor/Desktop/fanyi/models/inpainting/anime-manga-big-lama.pt
2025-05-27 01:17:03.824 | INFO | modules.inpainting.base:call:83 - hd_strategy: Resize
2025-05-27 01:17:03.826 | INFO | modules.inpainting.base:call:108 - Run resize strategy, origin size: (1859, 1400, 3) forward size: (960, 723, 3)
2025-05-27 01:17:03.827 | INFO | modules.inpainting.base:_pad_forward:61 - final forward pad size: (960, 728, 3)
GPTTranslation translation error: 'GPTTranslation' object has no attribute 'encode_image'
2025-05-27 09:35:46.793 | INFO | modules.inpainting.base:call:83 - hd_strategy: Resize
2025-05-27 09:35:46.794 | INFO | modules.inpainting.base:call:108 - Run resize strategy, origin size: (1859, 1400, 3) forward size: (960, 723, 3)
2025-05-27 09:35:46.795 | INFO | modules.inpainting.base:_pad_forward:61 - final forward pad size: (960, 728, 3)
GPTTranslation translation error: 'GPTTranslation' object has no attribute 'encode_image'

Here are the changed parts in base.py

user_prompt = f"""{extra_context}

你是一位资深漫画翻译，请将下方法语对白翻译成自然、口语化、富有情绪和人物性格的中文对白。
不要生硬直译，要像人物自己说话那样真实、有情绪。

原文：
{entire_raw_text}
"""

    entire_translated_text = self._perform_translation(user_prompt, system_prompt, image)

    # 🔽 把每一行翻译结果分配回文本框
    translated_lines = entire_translated_text.strip().split('\n')
    for blk, line in zip(blk_list, translated_lines):
        blk.text = line.strip()

[hitchkin]

You did not change the system prompt. The system prompt itself is found in lines 126 -132 (modules/translation/base.py)
and starts with
return f"""You are an expert translator who translates {source_lang} to {target_lang}. You pay attention to style, formality, idioms, slang etc and try to convey it in the way a {target_lang} speaker would understand... etc.

Forgot to add this:

• OpenAI is a bit too broad, what is the specific model (i.e. GPT-4.1, 4o, o3,...?). And in this case you should maybe give Gemma 3 a try, or Claude Opus 4 (through the custom option)
• You can increase the temperature in the settings which would allow for more creative/liberal translations.
• Only make those or any other changes - especially to the code - only if you are familiar enough with python/programming in general.

And this is more a fancy idea than a suggestion to be taken seriously:

• If you are really invested in this and have enough resources (time, ML knowledge, money, enough high-quality data -> in your case the french and chinese version of novels) you can absolutely fine-tune a model yourself using the novels that interest you to create a dataset for the fine-tuning. Whether this makes sense cost–benefit wise, is a question you'll have to answer for yourself.