段落摘要產生器
把一段文字濃縮成摘要、要點或關鍵句,保留原文的論點、證據和數字。
看一個例子
同一段文字,三種長度。下面是 Yila 對範例文字的真實輸出,每個數字都留在原處。
Artificial intelligence has entered clinical decision support faster than the frameworks meant to govern it. This study followed the introduction of machine-learning triage models at 47 hospitals over three years. Diagnostic accuracy rose by 23% against the conventional pathway and the median time to a working diagnosis fell by 34%, with the largest gains in emergency departments that paired the model with a structured handover. The gains were uneven: performance dropped for patient groups that were rare in the training data, two sites reported privacy incidents during data integration, and no site had a validation protocol for model updates. Clinician trust, measured by a quarterly survey, rose from 42% to 71% once the interface explained which findings drove each recommendation. The authors conclude that explainability and a routine validation cycle, not raw accuracy, decide whether such systems are adopted, and that regulation has not yet caught up with the pace of deployment.
At 47 hospitals over three years, machine-learning triage raised diagnostic accuracy by 23% and cut median time to diagnosis by 34%, but gains were uneven. The authors argue explainability and routine validation, not raw accuracy, decide adoption.
約為原文的 24%
A three-year study of 47 hospitals found machine-learning triage models raised diagnostic accuracy by 23% and cut median diagnosis time by 34%. Gains were uneven: performance dropped for patients rare in training data, and no site had a validation protocol for updates. Clinician trust rose from 42% to 71% once the interface explained which findings drove recommendations. The authors conclude that explainability and routine validation, not raw accuracy, decide adoption, and regulation lags deployment.
約為原文的 49%
This study followed machine-learning triage models at 47 hospitals over three years. Diagnostic accuracy rose 23% and median time to a working diagnosis fell 34%, with the largest gains in emergency departments using a structured handover. Gains were uneven: performance dropped for patient groups rare in training data, two sites reported privacy incidents during data integration, and no site had a validation protocol for model updates. Quarterly surveys showed clinician trust rising from 42% to 71% once the interface explained which findings drove each recommendation. The authors conclude that explainability and a routine validation cycle, not raw accuracy, decide adoption, and that regulation has not caught up with deployment.
約為原文的 72%
段落摘要產生器做什麼
段落摘要產生器讀一段文字,寫出一個更短的版本,保留這段文字的論點、支持論點的證據和關鍵數字。好的摘要是改寫式的:用更少的字重述論證,而不是隨意刪句子。「關鍵句」模式是例外:它原樣回傳原文中承載論證的句子,供你需要引用而不是轉述的時候用。
Yila 用工作台裡讀論文的同一個研究 Agent 來做摘要。它保留原文的術語、單位和數字,不加入原文沒有說的內容,並依原文的語言寫。每次摘要都是一個普通任務:它出現在你的歷史紀錄裡,你可以就這段文字繼續追問,也可以不用重新貼上就把它變成文獻筆記、一頁投影片或一篇更長的綜述。
怎麼用
- 貼上文字
- 貼上一段話、一篇摘要、一節結果或任何最長 5,000 詞的文字。計數器顯示你用掉了多少額度。
- 選長度和格式
- 短、中、長;輸出為連貫的摘要、要點列表,或原文的關鍵句。預設是中等長度的摘要。
- 在工作台裡讀
- 點「摘要」會在工作台裡開啟一個任務。摘要幾秒內出現,保存在歷史紀錄裡,你可以在同一個對話裡繼續就這段文字提問。
長度和格式
長度決定保留原文的多少,格式決定回答的形態。任何長度都可以搭配任何格式。
長度
- 短
- 一兩句話,不超過原文的五分之一:論點和主要結果。
- 中
- 一段三到五句話,不超過原文的三分之一:論點、證據、限制和結論。
- 長
- 不超過原文的一半,保留次要結果和限制。
格式
- 摘要
- 連貫的段落,適合閱讀或放進你自己的寫作。
- 要點列表
- 三到七條,每條一個發現,適合筆記和投影片。
- 關鍵句
- 原文中承載論證的句子,原樣引用。
研究者在哪些情境用它
- 把檢索結果裡每篇論文的摘要壓成一行,再決定讀哪些全文。
- 把密集的方法或結果段落壓成 journal club 需要的三句話。
- 準備引用某篇相關研究時,先抽出它的關鍵句。
- 自己的草稿超出字數限制時,壓短段落而不丟數字。
- 摘要一段外語文字,並用那種語言讀摘要。
- 回覆審稿意見前,先把每條長評論壓成一句話。
常見問題
一次摘要多少錢?
依模型實際消耗的 token 扣點數:一段短文約 3 點,5,000 詞的長文最多約 30 點。任務執行時會先凍結一小筆點數,結束後依實際用量結算。新帳號附有免費點數,最初幾次摘要不用額外花錢。
需要帳號嗎?
需要。摘要作為一個任務在你的 Yila 工作台裡執行,計費和保存都在那裡。用 Google 或電子郵件登入只需一步。
文字可以多長?
每次最長 5,000 詞。更長的文件更適合當檔案處理:把 PDF 上傳到「PDF 對話」,或者分節貼上。
會保存我的文字嗎?
貼上的文字和它的摘要會作為一個任務保存在你的帳號裡,和其他研究對話一樣。你隨時可以刪除這個任務,它也不會被用來訓練模型。
支援哪些語言?
摘要依你貼上的文字的語言寫:中文、英文、日文、韓文、西班牙文、德文、法文、葡萄牙文等。想換一種語言,在同一個對話裡說一句即可。
能摘要整篇論文嗎?
它摘要的是段落。整篇論文請用「PDF 對話」,它會讀完檔案並附頁碼回答問題;論文精讀的 Playbook 則會產出結構化筆記。
摘要準確嗎?
Agent 被要求保留原文的術語、數字和論點,不加入原文沒有說的內容;關鍵句模式原樣引用原文。像對待任何摘要一樣,依賴它之前請對照原文讀一遍。
