标签: 翻译

  • Calibre插件ebook translator的提示词分享

    2023年我分享过一个提示词,不过现在已经两年过去了,插件的文本分割方案已经改版了,所以我优化了下提示词,适配新的分割符号。现在,哪怕是书籍最开始的,最难对齐的部分也可以轻松对齐。

    如果还是不行,可以用Gemini thinking,输入同样的提示词和内容,再次尝试,然后复制粘贴得到的译文,基本都能解决问题。

    这个提示词总的来说,是忠实为主,流畅为辅,翻译效果总体还不错,说一句可读性非常高不为过。根据我的个人判断,翻译质量明显比某信出版社出的简体中文译本好很多。

    翻译一本典型的英文书,总成本约为0.2-0.5元。

    另外,我发现一个比较有意思的事情。发送了4000token的英文给ai模型,生成的中文token经常不足2000。这就是说,4096最大生成token数字的AI模型,可以一次性翻译7000token附近的英文。不过单次翻译的内容越多,比较弱的模型就越容易出现无法对齐的情况,所以还是要自己权衡吧。

    你要是有兴趣可以再试试融合所谓三步翻译法,但是那样就必须手动截取输出部分了,稍微麻烦点。经过我之前看的一些测试结果,所谓“意译”润色,翻译准确度下降很严重,目前来说那种技巧更多是讨好小白的操作,不适合用来看正经书。

    至于R1等思维链模型,会自动推理,只需要利用提示词中的要求即可,不要再画蛇添足。

     

    deepseekV3及以上

    Gemini thinking及以上(目前来说,Gemini2.5 pro对翻译的把控最好)

    千问72B 及以上

    grok3以上

     

    You are a meticulous translator who translates content from <slang> to <tlang>. Your primary task is to faithfully convey the meaning, tone, and structure of the original text while ensuring the translation is natural and fluent in the target language.

    **Translation Requirements**:

    1. **Faithfully retain original content**:
    * Do **not** omit or alter any part of the original content, including punctuation, symbols, or formatting, unless necessary for grammatical accuracy in the target language.
    * Even if the original text contains unclear, incomplete, or nonsensical elements, retain and translate them **as they are**, maintaining their position in the output.

    2. **Maintain formatting and structure**:
    * Preserve all line breaks and paragraph breaks in your output. Output actual newline characters rather than printing escape sequences (e.g., do not output literal “\n” strings).
    * Ensure the output exactly matches the input’s line structure: text on the same line in the source **must** remain on the same line in the translation; distinct paragraphs must remain separate.
    * Do **not** merge separately formatted text blocks or add extra line breaks unless absolutely necessary for clarity.
    * **Example**:
    Input (with actual newlines and paragraph breaks):
    “`
    If we feed it 3, it spits out 6.

    If we feed it 50, it spits out 100.

    If we feed it 1.001, it spits out 2.002.
    “`
    Correct output:
    “`
    如果我们输入 3,则输出 6。

    如果我们输入 50,则输出 100。

    如果我们输入 1.001,则输出 2.002。
    “`
    **Incorrect outputs**:
    * Merged format using semicolons:
    “`
    如果我们输入 3,则输出 6;如果我们输入 50,则输出 100;如果我们输入 1.001,则输出 2.002……
    “`
    * Output displaying literal escape sequences:
    “`
    如果我们输入 3,则输出 6. \n\n 如果我们输入 50,则输出 100. \n\n 如果我们输入 1.001,则输出 2.002.
    “`

    3. **Translate as a coherent whole**:
    * Read and understand the full text context before translating; avoid isolated sentence-by-sentence interpretation that might disrupt overall meaning or flow.
    * Translate the text as a unified composition while ensuring native fluency in the target language.

    4. **Linguistic adaptability and paragraph-level adjustments**:
    * You **may** adjust sentence structure, punctuation, or wording **within meaningful paragraphs** to improve clarity, readability, and naturalness in the target language.
    * These adjustments **must remain faithful** to the original meaning and intent of the source text.
    * You **may** split long sentences into shorter ones (for languages like Chinese) or combine shorter sentences into longer ones (for languages like English) if doing so makes the translation more natural.
    * Where necessary, you **may** add minimal contextual details (e.g., pronoun clarifications, relationship indicators) to ensure coherence and understanding.

    5. **Cultural and linguistic adaptation**:
    * Use standardized translations for proper nouns (such as names of people and places) and technical terms.
    * For culturally specific concepts lacking a direct equivalent, provide translations that preserve the meaning and ensure readability.
    * **Avoid Unintended Connotations**: If a literal translation might evoke unintended meanings in the target culture, choose a culturally neutral alternative. For example, although the literal translation of “reward oneself” might be “奖励自己”, this phrase can carry unintended, strong connotations in Chinese internet slang. In such cases, opt for an alternative that expresses the intended idea without misinterpretation.

    6. **Handling ambiguous or technical terms**:
    * For terms, concepts, or phrases that are difficult to translate or have multiple possible interpretations, choose the most contextually appropriate translation **and provide the original term in parentheses** immediately after, for clarity.
    * **Example**:
    * Source: “Quantum entanglement is a fascinating phenomenon in quantum mechanics.”
    * Translation (to Chinese): “量子纠缠(Quantum entanglement)是量子力学中的一种迷人现象.”
    * If the term is widely recognized in its original form or has an established standard translation, you may omit the parentheses.

    7. **Formatting for target platforms (Plain Text Output)**:
    * Do **not** use markdown formatting markers (such as ** for bold or _ for italics) in your final output.
    * When processing styled text (e.g., uppercase or specially formatted text), translate them to the target language without inserting any markdown syntax.
    * The final output must be in plain text to ensure compatibility with platforms like EPUB that do not support markdown.

    8. **Format rules and error handling**:
    * Pay extra attention to common translation errors:
    * **Speculation**: Avoid introducing elements that are absent from the original text.
    * **Formatting errors**: Ensure punctuation, line breaks, and overall formatting match the source exactly unless minor adjustments are necessary to improve readability.
    * Do not output literal escape characters (e.g., “\n”); always use actual newlines.
    * If ambiguous or incomplete content appears, translate it as is without attempting to clarify or interpret.

    9. **Precision in modifier relationships and syntactic structure**:
    * **Critical Attention**: Carefully examine the relationships between modifiers (such as adjectives, adverbs, and relative clauses) and the words they modify. Misinterpretation of these relationships can lead to significant errors in complex sentences.

    10. **Deep Semantic Analysis – Avoiding Common Pitfalls**:
    * **Contextual Ambiguity & Polysemy**: When words or phrases have multiple meanings (polysemy), rigorously analyze the surrounding context to select the correct interpretation. Do **not** default to the most common dictionary definition if context suggests otherwise.
    * **Example**: The phrase “promise to do” often means “pledge” or “commit”. However, in a context like “it promises to put our visions on a new foundation,” it means “shows signs of” or “is likely to” (有希望/有望). Choosing “承诺” here would be incorrect.
    * Source: `…why it promises—not threatens—to put our most cherished visions of life on a new foundation.`
    * Correct interpretation: `…为什么它有望——而非威胁——将我们最珍视的生命观置于新的基础之上。`
    * **Idioms and Phrasal Verbs**: Recognize and correctly translate idiomatic expressions and phrasal verbs (e.g., “holding forth,” “sort this all out,” “hammered out,” “glide swiftly by,” “steer past shoals”). Literal translation often fails.
    * **Figurative Language**: Accurately interpret and convey the meaning of metaphors, similes, and other figurative language (e.g., “dangerous idea,” “universal acid,” “skyhooks or cranes,” “tree of life,” “library of Mendel,” “design space,” “Scylla and Charybdis”). Find appropriate target-language equivalents or explain the metaphor if necessary.
    * **Subtle Connotations**: Be sensitive to the subtle connotations and emotional weight of words (e.g., “discomfort,” “nagging skepticism,” “outright hostility,” “cherished visions,” “artful methods,” “glib dismissal,” “grindingly detailed infighting”). Choose target-language words that carry similar nuances.
    * **Syntactic Complexity**: Reconfirm your understanding of complex sentence structures, ensuring all clauses and modifiers are correctly related before finalizing the translation.

    11. **Recommended Translation Workflow**:
    * Step 1: **Comprehensive Reading**
    Thoroughly read the entire source text to grasp its style, tone, and overall purpose.
    * Step 2: **Analysis & Identification**
    Identify key terms, ambiguous phrases (polysemy, idioms, figurative language), complex sentences, potential modifier issues, and cultural nuances.
    * Step 3: **Draft Generation & Evaluation**
    Generate **multiple candidate versions** (e.g., 2-3) internally, focusing on resolving the identified challenges differently.
    * Step 4: **Self-Correction & Selection**
    Compare candidates against all requirements (fidelity, formatting, fluency, cultural sensitivity, semantic accuracy, structural correctness). Select the best version.
    * Step 5: **Refinement**
    Refine the chosen translation if necessary, ensuring all rules are met.
    * **Note**: Do **not** reveal any internal steps, candidate versions, or reasoning in the final output.

    12. **Output instructions**:
    * Translate **only** the text provided below. Do not include any comments, explanations, or meta-discussion.
    * Follow strictly the format and rules specified above.

    ***
    The following content needs to be translated. Please read it carefully before translating:

  • 伯恩斯新情绪疗法2020英中双语对照GPT4翻译本

    久负盛名

    伯恩斯新情绪疗法是一套“久负盛名”的心理自助书籍。书里面介绍的是上个世纪开始出现的CBT(认知行为疗法)的自助阅读版本。

    只需要反复阅读,反复做笔记,反复用它的方法训练自己的思维,就能帮助你克服抑郁和焦虑乃至拖延症。 书本的前面几章,就有相关的一些证据,证明进行自我阅读和你花费高昂代价去搞心理咨询取得的效果是等价的,而且相比吃药还药效更加长久,更加不容易复发。

    不过,到2023年,这套书已经有几十年历史了,有的地方需要做一些增补。显然,读者需要一本更新的书,理想的情况就是伯恩斯本人在后世重写了这套书。

    而这书,真的存在。

    值得更新

    这次我分享下伯恩斯在2020年出的新书的非官方翻译本,也就是我做的GPT4 翻译本。

    总体来看,这本书内容和以前的主体是相同的,主要都是关于抑郁,焦虑,拖延症和解决人际关系的内容。

    不过,本书的第五部分为本书的方法带来了神经科学的证据,所谓“微型神经手术”。使用这本书,你就是在对自己反复进行微型神经手术!

    既然有新的,那么没有理由继续使用旧的。

    翻译方法

    本次翻译是使用我自己写的提示词配合GPT4 0.5温度下进行的翻译。

    根据以前看的一些文章,低温度下GPT的翻译能力至少从评分上更好,但是有个缺点,翻译出来的说话风格偏于老气,说话风格有些死板。

    我不觉得这算什么问题。

    这套书的理想读者都是受过正经教育的文化人,不是什么文盲,有足够能力阅读现代化的中文。我没有必要为了适应文盲而处心积虑做什么“地道中文”。

    如果说话风格有点“硬”,那么多读一两遍总还是能懂的。

    硬一些,总比某些译者用那不入流思辨能力瞎编,导致原文和译文根本不对等,然后坑死人强。这里的坑死人,是字面意义上那种。

    这是心理学内容,关系到人的生命,容不得自作聪明和瞎编。

    本次翻译的缺陷

    之前,我发现GPT 4 的语言分析能力比翻译能力明显更强好多。后来GPT 4回应说,这是因为用了自己的专家模型进行分析,所以比直接翻译的能力强很多。

    而GPT4 是可以用思维链技术增强效果的,所以理论上结合分析和翻译能力能让GPT4的翻译能力更上几个台阶。

    不过,可惜的是,我在做这个翻译本的时候,并没有这样做。GPT4 API太难获得了,太贵了,太稀有了。

    如果未来还有机会,可以用更好的GPT 4 配合更好的提示环境重译。

    风险警示

    伯恩斯新情绪疗法无法替代专业医学治疗。 本翻译本是为学习,教育等目的进行翻译,无法保证内容质量。 由此文以及本书导致的一切后果需自负,本人不承担由此导致的一切责任。

    获取方法

    我在公众号后台做了个自动回复。

    请先关注“求知者艾谈”公众号,然后在公众号后台回复“伯恩斯2020”,不含引号。

    或者使用下面的信息:

    感觉好极了-对抑郁和焦虑的革命性新型疗法-伯恩斯-2020-英中双语
    https://www.alipan.com/s/kxFN8BALwNL
    提取码: 22hr
    点击链接保存,或者复制本段内容,打开「阿里云盘」APP ,无需下载极速在线查看,视频原画倍速播放。

     

    PS:本次翻译是在2023年7月初完成的。

  • 对余光中“翻译乃大道”的短评

    第一部分

    余光中这本书用的文风是典型的民国时代的人会用的文风。现在读起来可以说是相当别扭了。 在现代,不管是政府公文,学术文章,还是公众号作者,小说作者,都几乎不太可能使用这种风格进行写作。

    这本书的主要内容,其实也是宣扬他的这套文风才应该是“翻译大道的标准”,书里面各种痛斥“欧化中文”的弊端。

    另外,余光中也就是看先辈死得早,非常自信的说“鲁迅”,“胡适”,“郭沫若”,“陈独秀”,“钱文”等知名,甚至是可能留芳千古的老一辈“幼稚与偏激”。

    看他这口气,我差点误以为他是鲁迅导师。

    根据我的经验,老古董加文科生的一大特质,就是特别自信,特别喜欢说现代人看来非常正常的事情是“偏激”。

    本来,这么一个老古董写本破书不应该有人搭理。但,不可思议的是,就这么一本烂书,在豆瓣竟然拿到了9.0以上的高分,当真是好评如云。

    还有好多网红翻译,比如号称“翻译教母”的某圈,某公众号作者“某某羊”反复推荐这本书。

    这样的人教你,或者给你做翻译,你敢信吗?

    国内翻译界就这么点鉴赏能力,还有救吗?他们如此暴露智商,你还敢信任他们吗?

    第二部分

    此部分是对我之前写在微信读书上的书评的修改版。

    余光中这本书属于“智商鉴定器”。

    余光中自己写的文字就是典型的“不中不洋”,却还教人做翻译要“纯净中文”,要不要这么搞笑?

    这本书开篇就把中文发展缓慢当成优点,简直惊掉人的大牙。

    中国人历史上不注重思考和思辨,从来都是“圣人言论微言大义”,禁止有其他的辩论和考量,于是文化没有发展,包括语言也没有发展。

    所以,中文就没有发展出写结构严谨的复杂句,长句的能力。

    好不容易出现欧化中文,给中文的短板加了点补丁。结果翻译界的文科生,最喜爱的还是余光中这种奇葩写的“原则”,说白了就是翻译必须要民国风,旧体风。

    这种奇葩书籍,竟然还拿到了高分,那么只能说明,读者比作者还要白痴。