M-fabrik 各項工具的上線與更新記錄，新在上。只記對使用者有意義的變動——上線、能力提升、語料入庫；內部實驗與工程細節不在此列。 Release notes for the M-fabrik tools, newest first. Only changes that matter to users are listed — launches, capability gains, corpora added; internal experiments and engineering details are not.
詞典來源說明：《滿漢大辭典》《新滿漢大詞典》《満漢和辭典》三部對接日本東北大學東北アジア研究センター的滿語辭典網站（hkuri.cneas.tohoku.ac.jp）；語料詞表對接 manc.hu；人名詞典對接香港金由美教授（Loretta Kim，LEK Research Collective）的 NCNHNL 資料庫。各家資料的著作權歸原作者與機構所有，本站僅作檢索之用。 Dictionary sources: the three dictionaries 滿漢大辭典, 新滿漢大詞典 and 満漢和辭典 are accessed through the Manchu dictionary site of Tohoku University (Japan)'s Center for Northeast Asian Studies (hkuri.cneas.tohoku.ac.jp); the corpus vocabulary through manc.hu; the name lexicon through the NCNHNL database of Prof. Loretta Kim (LEK Research Collective, Hong Kong). Copyright of each dataset remains with its authors and institutions; this site only queries them.
轉寫結果下方改為完整的逐詞解析，並顯示詞典命中率；轉寫文字本身可直接點開改動，改完即以新文字重新解析——OCR 認錯一兩個字母時不必再回頭重跑整頁。另修一個顯示錯誤：譯文顏色原先寫死為近白色，深色主題下正常，淺色主題下卻與紙色相近、幾乎看不見；現已改為隨主題切換，兩種主題下都清晰。 Transcriptions now carry a full word-by-word analysis underneath, with a dictionary hit rate. The transcription itself is editable in place, and the analysis re-runs against your corrections — no need to re-run a whole page because OCR missed a letter or two. Also fixed: translated text was hard-coded to a near-white colour, which read fine in dark theme but was almost invisible against the paper background in light theme. It now follows the theme.
兩處數字此前站內不一致，現已統一到可複算的口徑。① 單位一律用「頁」：已發布數據每張掃描圖即一頁，故系統計數不再混用「葉」，「葉」只保留在具體書的版本著錄裡。順帶更正兩處實數——《大清會典》卷一的 27 項實為 27 個半葉，《滿文老檔》3,803 項為頁（原檔一頁一葉）。② 詞典原形數改報 65,196：舊值 66,509 是詞典檔的原始行數，把同一個詞的三種轉寫變體與無釋義的存根都各算了一條；新值為歸一化合併後的實際原形數（單詞 22,925 + 多詞 42,271）。 Two sets of figures disagreed with themselves across the site and have been reconciled to counts you can recompute. (1) One unit: the page. Every published scan is one page, so system-level counts no longer mix in “folio”; that term now survives only where a specific edition is being described. Two real corrections came out of it: the 27 items in juan 1 of the Collected Statutes are 27 half-folios, and the 3,803 items of the Manchu Veritable Records are pages (one to a folio in the original). (2) The dictionary now reports 65,196 base forms. The old 66,509 was the raw line count of the dictionary file, counting three transcription variants of the same word — and entries with no gloss — as separate items. The new figure is the normalised, merged total (22,925 single-word + 42,271 multi-word).
第 41–42 冊共 505 頁，取柏林國家圖書館的武英殿刻本與法國國家圖書館的寫本（Mandchou 144(6)，1770）並排互校。兩本逐頁對齊後標出 561 處分歧：疑 OCR 誤讀 136、疑版本異文 205、待判 220。分歧一律不自動合併——刻本與寫本性質不同，抹平差異等於銷毀校勘信息，頁面只並列呈現並標注類別。判據的局限也寫在頁上：詞典命中不等於此處正確。 505 pages from fascicles 41–42, setting the Wuyingdian printed edition (Berlin State Library) beside the manuscript (BnF, Mandchou 144(6), 1770). After page-level alignment the two witnesses differ in 561 places: 136 likely OCR errors, 205 likely genuine textual variants, 220 undecided. Nothing is merged automatically. A printed edition and a manuscript are different kinds of witness; flattening their differences would destroy the collation itself, so the page only juxtaposes and labels. The limits of the classifier are stated on the page too: a dictionary hit does not mean the reading is right here.
① 《滿文老檔》太宗朝 2,600 頁補上掃描原圖。此前引用太宗朝內容時縮圖是破的——並非顯示故障，而是這 2,600 頁的影像欄從一開始就是空的（太祖朝 1,203 頁一直齊全）。現已全部補齊，兩站各 3,803 頁。同時加了防禦：日後若再有無影像的數據，會顯示「暫無影像」占位而不是破圖。② 英文提問改用英文作答。原先無論用什麼語言提問都回繁體中文。現在作答語言跟隨提問語言，但檢索關鍵詞仍一律用中文／滿文轉寫——因為老檔本身沒有英文欄位，拿英文詞去檢索會落空。 (1) 2,600 Taizong-reign pages of the Manchu Veritable Records now have their scans. Thumbnails for that reign were previously broken — not a display fault, but an image field that had been empty from the start (the 1,203 Taizong-era pages were always complete). All 3,803 pages now carry images on both servers. A guard was added as well: data without an image now renders a “no scan” placeholder rather than a broken thumbnail. (2) Questions asked in English are now answered in English. Answers previously came back in Chinese whatever the question’s language. The answer now follows the question — while search terms are still issued in Chinese or Manchu transcription, since the records carry no English field and English keywords simply find nothing.
v0.5 曾為了讓譯文中英雙出、標註殘缺字而拿掉了範例，語域也跟著飄。這次只改一個變量——按體裁（奏對／故事／歷史）各給三則範例作「語域參考」，其餘一律照 v0.5。96 句盲評（逐句隨機、打分前不揭盲）：語義充分性由 3.34 升至 3.52（滿分 4），意思正確率 79.2% → 86.5%，低分句由 20 句減至 13 句；歷史類提升最大（3.35 → 3.82）。但增益未達統計顯著——96 句的自助法信賴區間下界壓在 0，只能說「一致有利於 v0.6」，不能說「顯著更好」。意外的收穫在別處：範例真正壓住的是實詞錯譯而非文體——hiya（旱）此前被譯成「侍衛」、hangkame（口渴）被譯成「浸濕」（意思相反），人名地名「汝寧→朱寧」「南郭→楠哥」，這幾類在 v0.6 都改對了；合理的解釋是範例裡出現了真實的人名地名與公文用詞，給了模型正確的候選先驗。另需說明：字符級指標 chrF++ 雖由 18.85 回升到 21.62，仍未追平更早的 v0.3（24.83），且該指標會因文體詳略差異而誤導，不宜單獨用來判定版本優劣。「實驗性」標示維持不變——3.52 分與 86.5% 尚不足以撤下，且奏摺類目前沒有人工參考譯文，其品質仍無定量證據。 v0.5 had dropped its examples in order to emit Chinese and English together and flag illegible graphs, and register drifted as a result. This change moves one variable only: three examples per genre (memorial dialogue / narrative / historical prose) are restored as a “register reference”, everything else as in v0.5. Blind evaluation over 96 sentences (randomised per sentence, unblinded only after scoring): semantic adequacy rose from 3.34 to 3.52 out of 4, correct-sense rate from 79.2% to 86.5%, and poorly-rendered sentences fell from 20 to 13, with historical prose gaining most (3.35 → 3.82). The gain is not statistically significant: with 96 sentences the bootstrap confidence interval touches zero at the lower bound, so this is “consistently in v0.6’s favour”, not “significantly better”. The real find lay elsewhere. What the examples suppressed was mistranslated content words rather than register: hiya (drought) had been rendered “imperial guard”, hangkame (thirsty) as “soaked” — the opposite sense — and place and personal names came out wrong (Runing as “Zhuning”, Nanguo as “Nange”). All of these are correct in v0.6, plausibly because the examples supplied real names and administrative vocabulary as priors. One caveat: the character-level chrF++ score recovered from 18.85 to 21.62 but still trails the earlier v0.3 (24.83), and that metric is misled by differences in register and verbosity, so it should not decide a version on its own. The “experimental” label stays. 3.52 and 86.5% are not enough to remove it, and for memorials there is still no human reference translation and therefore no quantitative evidence at all.
體例照美國研究生院的 Reading for German／French：純語法、受控詞彙、真實原文精讀，不做發音與口語。不從字母起步——字形只作認讀附錄，全書以拉丁轉寫直取語法核心。20 課語法主線、6 篇《百條》會話精讀（註釋逐篇遞減至無註裸讀）、4 種附錄速查表，另附 29 個 Anki 卡組共 1805 張卡，每課一個檔案。語法骨架與全部原文取自 Möllendorff 1892（公有領域）。本頁需口令。 Modelled on the American graduate “Reading for German / French” courses: grammar, controlled vocabulary and close reading only — no pronunciation or speaking. It does not start from the script: letterforms are a recognition appendix, and the body of the course works in romanised transcription, going straight for the grammar. 20 grammar lessons, 6 annotated dialogues from the Tanggū meyen with scaffolding withdrawn lesson by lesson, 4 reference appendices, plus 29 Anki decks (1805 cards), one per lesson. Grammar and all texts from Möllendorff 1892 (public domain). Password required.
解析結果與 OCR 轉寫下方新增「朗讀」按鈕，可直接聽整句。滿語已無母語者留下的可用語音資料，聲學模型借用近親語言的聲庫，發音僅供辨識輔助，不作標準音依據。 A “read aloud” button now sits under both parser output and OCR transcriptions. No usable native Manchu speech corpus survives, so the acoustic model borrows a related language’s voice: treat the audio as a recognition aid, not as a pronunciation standard.
柏林國家圖書館滿文佛經 1910 頁全數取回並校驗；法鼓佛教學院滿文大藏經（CC BY-SA 3.0）光碟全量下載，含 108 函 781 條滿／漢／藏三語經目與 996 條滿-漢-梵-蒙四語佛教術語。兩批合起來補上了 18 世紀後期刻本的空缺。 1910 pages of Manchu Buddhist scripture from the Staatsbibliothek zu Berlin, fetched and checksummed; plus the complete DILA Manchu Canon disc (CC BY-SA 3.0) with a 781-entry Manchu/Chinese/Tibetan catalogue across 108 fascicles and 996 four-language Buddhist terms. Together they fill the late-18th-century gap in our holdings.
舊的對比曲線會把淡黃墨的字連同背景一起推白，律例類刻本受害最重。新做法按頁自適應：先選對比最強的顏色通道，曲線中點取墨與紙的中值（筆畫粗細不變、不粘連），斜率按該頁墨紙落差自動定。專家批改集逐頁對比，新法 21 勝 3 負 1 平。 The old contrast curve pushed faint yellow-brown ink white along with the background, worst of all on legal-code editions. The new pass adapts per page: pick the channel with the strongest contrast, anchor the curve’s midpoint between ink and paper (so stroke weight is preserved and glyphs don’t merge), and set the slope from that page’s ink-to-paper gap. Against an expert-corrected set the new pass wins 21 pages, loses 3, ties 1.
全書逐頁預處理後過 OCR，結果進專家批改台等待校訂。同時把預處理獨立成通用模組，其他刻本可直接複用。 Every page preprocessed and run through OCR, with the output queued in the expert correction desk. The preprocessing was also split out as a reusable module for other printed editions.
原先分散的幾個工具頁合成一個品牌站：首頁即滿語解析、另有 OCR、數據庫示例與使用說明，統一雙語頁眉、深淺主題與繁體介面。隨後一週把三頁重寫並讓主題設定跨頁保持。 The scattered tool pages became one site: language query on the front page, plus OCR, database samples and a guide, with a shared bilingual header, light/dark themes and a traditional-character interface. Over the following week the three pages were rewritten and the theme choice made to persist across them.
可用中文（或中滿混輸）反查滿文詞。整句自動分詞，複合詞先給整體、再拆構件；命中以三色標示（綠＝命中、黃＝同音近似、紅＝未收）；音譯人名地名不再被拆散（如「乾清宮」保持整體）。同時把金典人名系統立為第 6 部並聯字典。 Look Manchu words up from Chinese, or from mixed Chinese-Manchu input. Sentences are segmented automatically; compounds are shown whole first, then decomposed; hits are colour-coded (green = found, amber = near-homophone, red = not attested); transliterated names stay intact rather than being split. A sixth dictionary — Prof. Kim’s name system — was added to the parallel lookup at the same time.
合併詞典此前對多義詞只留一個義項，遇到 somimbi 這類詞會丟掉主要義。現在核心義項至少保留兩個，翻譯改為按上下文擇義。另修掉一批因暗寫 š／ū 而查不到的詞。 The merged dictionary used to keep a single sense per entry, which lost the primary meaning of words like somimbi. Core entries now retain at least two senses, and translation picks a sense by context. A batch of words unreachable because of unmarked š/ū was fixed as well.
字典開始記版本號。v1.1 併入金教授的滿漢人名系統，其中 11847 條帶轉寫者直接可查。讀檔案時人名不再是黑洞。 The dictionary now carries version numbers. v1.1 merges Prof. Kim’s Manchu-Chinese name system; the 11,847 entries with romanisation are directly searchable. Personal names in archival documents are no longer a blind spot.
輸入一個滿文詞形，給出切分、還原到詞典形（-mbi）、八格與詞類判定，並列出各部字典的原文詞條。支援滿文 Unicode 直接輸入。詞形覆蓋率由 82% 提到 97%。 Give it an inflected Manchu form and it returns the segmentation, the dictionary form (-mbi), case and part-of-speech, plus the raw entries from each dictionary. Manchu Unicode input works directly. Form coverage rose from 82% to 97%.
885 件文書完成轉寫、雙語翻譯與提要，即為「數據庫示例」頁展示的那一批。 885 documents transcribed, translated into two languages and summarised — the set shown on the Database samples page.
上傳一頁滿文影像即得轉寫，隨後接入滿→繁中／英翻譯模組。服務端自帶預處理，手機拍攝的照片也能用。 Upload a page image and get the transcription; a Manchu-to-Chinese/English translation module was wired in shortly after. Server-side preprocessing means phone photographs work too.
印刷體滿文的字元錯誤率首次低於 1%，即約每 110 個字母錯一個。奏摺與草書體另有專門模型，錯誤率高於此數。 Character error rate on printed Manchu fell below one per cent for the first time — roughly one slip in every 110 letters. Memorials and cursive hands are served by separate models and remain less accurate.
同一批頁面、同一口徑，逐頁列出我們與既有系統的轉寫結果，供讀者自行判斷。 Same pages, same measurement: our transcription next to the existing system’s, page by page, for readers to judge.