fanlusky
c064cc641a
问题:merge_alternative_segmentations 把首选读音与所有备选读音的候选
混在一起、纯按词典词频排序(注释自陈"跨切分词频不可直接比较",实现
却照排)。于是输入 gau 时,高频的 改/该(gai，邻键，权重13)盖过了
低频的 挂(gua，换位，权重10),尽管 gua 的编辑代价更低、更贴近"交换
字母顺序"的用户意图。
改动:
- AutocorrectCut 增加 edge_count/weight,report k-best 的编辑代价;
same_cost_as() 判定同代价。
- resolve_series_query 按代价把备选拆成两层:与首选同代价的进入词频
竞争(alternative_corrected_cuts,保留 uanli→{quan,cuan,...} 那类
等权消歧),严格更贵的落入 costlier_corrected_cuts。
- query_exact 将更贵读音排在整个首选代价层之后追加(仍查询、去重、
标注 corrected_from),词频不再能把更贵的纠错顶到便宜的前面。
权衡(重要):离线 eval(seed 42,300 样本)显示这是"编辑代价优先"
对"词频优先"的取舍——mixed 模型 R@1/R@3 各档下降约 8~17 个点,
deletion 模型小幅下降 1~3 点。原因是注错样本的原词分布跟随词频,
词频优先在整体召回上更强;本改动以召回换取"最便宜纠错优先"的确定性
排序,是产品取向的选择,经维护者确认后合入。
验证:imetest 全绿(新增 gau 代价分层用例、元音尾守卫用例);
scripts/format.sh --check 通过。
Metasequoia IME for Windows
中文 README · Website · Docs · Privacy
A free and open-source Chinese and Japanese input method for Windows 10 and 11, built on TSF with no legacy IMM32 path.
This repository is the whole Windows product: the TSF text service DLL, the resident server, the conversion engine, a self-written native GUI framework, the HTML surfaces and the installer. The engine was imported from MSIME-Engine and specialized for Windows; it now lives in engine/ as a first-party component, so a plain clone builds the product. See engine/UPSTREAM.md.
This is a public beta. Expect rough edges and read the release notes before upgrading.
Install
Download the installer from Releases or the download page, run it, then switch with Win + Space.
Current builds are not code signed. SmartScreen will block them and uiAccess is unavailable, which means the candidate window cannot float above elevated applications. Verify the download with the SHA256 published on the download page.
You need the Microsoft Visual C++ 2015–2022 Redistributable (x64). The server and settings app are 64-bit; the x86 redistributable is a separate package and does not substitute for it. If you cannot switch to the IME, the server keeps exiting, or the settings window closes immediately, install that first.
What it does
- Chinese input: full pinyin, double pinyin (Xiaohe / Ziranma / Shoudao / Microsoft), Wubi 86
- Japanese input: romaji, with hiragana, katakana and dictionary candidates; also available as a temporary mode from Chinese
- Helpcode (形码) filtering on pinyin schemes: five tables, single- and double-code
- Optional cloud candidates, and AI suggestions via DeepSeek / OpenAI / SiliconFlow / Groq
- English-Chinese glosses in the vertical candidate window
- Mixed Chinese-English input, emoji and kaomoji, a dedicated English candidate mode
- Voice input with streaming recognition and optional cleanup
- Handwriting, on-screen keyboard, floating toolbar, clipboard history
- User dictionary management, candidate frequency tuning, skins with light and dark variants
Privacy
An input method sees every keystroke, so the boundaries are stated explicitly in PRIVACY.md: which features reach the network, what each one sends, what the defaults are, and how to turn each off.
Two things worth knowing up front. Cloud candidates are on by default — the composition spelling goes to Google's input-tools service, and it is the only network feature that works without you supplying a credential. API tokens are stored in plain text in config.toml, unlike macOS and Linux which use the Keychain and Secret Service.
There is no telemetry, analytics or crash reporting of any kind.
Building and contributing
CONTRIBUTING.md has the environment, the build order and what to run before opening a pull request. Contributions do not have to be code — dictionary entries, documentation, icons and compatibility testing all count, and the recruiting page lists them by area.
Note that most issues, documentation and code comments are in Chinese. English pull requests and issues are welcome.
Licence
GPL-3.0. See LICENSE and THIRD_PARTY_NOTICES.txt.