AI 与机器学习2020arxiv.org
迈向全自动漫画翻译
Towards Fully Automated Manga Translation
进阶 · 约 45 分钟读完前置知识：机器翻译基础、计算机视觉基础、日语书写系统常识
导读
这篇论文提出一个全自动漫画翻译流水线，尝试将日文漫画端到端翻译成英文。系统不仅做文本机器翻译，还结合了分格布局、角色性别等图像信息来辅助翻译。作者认为这有望大幅降低翻译成本、让更多漫画被翻译出来。
HN 关注点：机器翻译能否胜任漫画这类强依赖图像语境和文字游戏的内容，讨论焦点是自动翻译的质量能否达到人工水平，以及它对翻译社区和盗版生态的影响。
HN 评论精选
这几条评论里：质疑 1 补充 1 亲历 1
质疑自动化翻译无法达到优秀人工翻译的质量，只对非授权翻译和监控技术有用
翻译质量对阅读体验影响巨大。我手头有四个以上不同译者的《孙子兵法》译本，还有但丁、伊利亚特、陀思妥耶夫斯基等——有些译本根本读不下去，只有一两个出色。自动化不可能产出真正好的翻译。这类技术唯一的用处是生成那些本来不会被翻译的作品的非授权译本，但这也不算什么大事，因为社区早就把几乎所有作品都汉化/英化了。补充：另一个用途是帮助构建能结合双模态信息流、理解肢体语言的爬虫，如果成功，最终可能会被用于监控摄像头、麦克风等自动监控技术。
hertzrat查看英文原文
The quality of a translation makes a dramatic difference in the reading experience. I have 4 or more copies of the art of war, each with different translators. Dante’s inferno, the Illiad, Dostoyevsky, Beowulf, Confucius: some translations are unreadable and one or two incredible. Automation is just not going to produce the really good translations. The only benefit of a technique like this will be to generate unlicensed translations of work that will not otherwise get translated, which isn’t even that big a deal because the communities already produce unlicensed translations of just about everything. Edit: The other use is to help create better context aware data scrapers that can combine bi-modal information streams and add some body language understanding. I guess it will probably end up in automated surveillance tech like security cameras/mics etc if it works.
补充机翻在漫画翻译圈不被接受，且现有技术忽略了排版、字体、振假名等大量语境信息
在翻译社区里，即使在漫画翻译流程中部分使用机器翻译也不被认可（想在有偿工作中这么做更是没门），连盗版网站上的汉化组也普遍反感机翻。优秀的机器翻译加质检编辑团队或许能收拾干净，但最终你可能还是直接雇一个双语流利的人更省时省钱。看到研究视角很有意思。当前翻译技术没有考虑分格布局、横竖排版、字体差异、手写与印刷体、假名汉字混用等作者用来传达语气、暗示说话人或添加潜台词的手段，还有英语文学中完全没有对应物的东西，比如用振假名给一个词附加第二种读法或双关含义。论文只迈出了几小步——输入了分格布局和角色性别等基本信息，期待后续发展。
kevingadd查看英文原文
In translation communities the idea of using automated translation even in part in the process of translating manga is frowned upon (good luck if you try to get away with doing this on a paid gig), and even the people pirating manga on websites tend to actively dislike automated translation as well. Heroic efforts from an excellent machine translation stack and a team of quality checkers and editors can help clean it up, but at the end of the day you'd probably save time and money by just hiring someone fluent in both languages. It's always interesting to see research perspectives on this. I think there's considerable room for improvement, especially given that current translation tech doesn't take into account things like panel layout, horizontal/vertical layout of text, different typeface use, hand-written vs. typeset text, and mixed katakana/hiragana/kanji usage - all things that an author may use to convey tone, implicitly identify the speaker, or add subtext. There are also things with no equivalent in English literature/comics whatsoever, like using furigana to attach a second reading (or dual meaning) to a word. I can imagine some of this info eventually being pulled in by built-for-purpose MTL tools if a research group puts in the time and energy to do it - the paper appears to be a couple brief steps in that direction because they're feeding in basic information about panel layout and the genders of the people in the panels, but little else. Excited to see what happens in this space even if the idea of even more people trying to translate dense comic prose with Google Translate fills me with dread. EDIT: I should have done more research before posting this... digging into the authors of this paper, two of them are working at a company that is trying to sell this current technology to authors right now despite the fact that it is not adequate for the challenges of translating comics to other languages. How depressing.
亲历亲历者：机翻可作为人工翻译的辅助查词工具，这项研究或能改进这一环节
我做过志愿漫画翻译（后来我们翻译组拿到资助后也开始有偿翻译），很大一部分工作是用谷歌翻译或翻译工具辅助查词、寻找翻译灵感，所以这项技术可能在这方面有帮助，省去使用谷歌翻译等工具的需要。
forkLding查看英文原文
I've been a volunteer manga translator in the past, (also started getting paid for it after a while when our manga translation group started getting funded) and a large part of it is using google translate or translation tools to help out with definitions or look for inspiration in translating the words in the manga so this could likely help out in that regard by saving the need of tools like google translate
相关资源
从这篇论文所有 HN 讨论帖的全部评论里整理出来的链接。
历次讨论
- 2020Hacker News 讨论帖144 分 · 75 评
同领域还值得读
DeepSeek-R1：通过强化学习激励 LLM 的推理能力
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL
研究如何用强化学习激励大语言模型的推理能力，提出 DeepSeek-R1，并探讨 RL 训练对模型推理表现的影响。
AI 与机器学习2025 · 1,351 分 · 1,056 评 · 引用 30进阶 · 约 80 分钟编辑精选1-bit LLM 时代：面向低成本计算的三值参数
The Era of 1-bit LLMs: ternary parameters for cost-effective computing
提出用三值参数 {-1, 0, 1} 替代 LLM 中的浮点权重，即所谓“1-bit LLM”。
AI 与机器学习2024 · 1,040 分 · 447 评 · 引用 19进阶 · 约 45 分钟编辑精选DeepSeek-V3.2：推进开放大语言模型的前沿
DeepSeek-v3.2: Pushing the frontier of open large language models
这是 DeepSeek-V3.2 的技术报告，延续了 DeepSeek 开源权重模型的路线，目标是让开放模型继续逼近甚至追平前沿闭源模型。
AI 与机器学习2025 · 982 分 · 465 评进阶 · 约 90 分钟编辑精选通过自我对弈与通用强化学习掌握国际象棋和将棋（AlphaZero）
Mastering Chess and Shogi by Self-Play with General Reinforcement Learning
DeepMind 的 AlphaZero 论文：用通用强化学习算法（自我博弈 + MCTS + 神经网络），从零开始仅数小时训练便在国际象棋和将棋上击败了世界顶级引擎 Stock…
AI 与机器学习2017 · 539 分 · 270 评 · 引用 13进阶 · 约 90 分钟编辑精选用强化学习发现更快的矩阵乘法算法
Discovering faster matrix multiplication algorithms with reinforcement learning
论文提出 AlphaTensor，用强化学习搜索矩阵乘法算法。
AI 与机器学习2022 · 518 分 · 111 评专业 · 约 90 分钟编辑精选