AI系統中的偏差與偏見

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"本文分享自百度開發者中心","attrs":{}},{"type":"link","attrs":{"href":"https://developer.baidu.com/article.html#/articleDetailPage?id=293562?from=010803","title":"","type":null},"content":[{"type":"text","text":"AI系統中的偏差與偏見","attrs":{}}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"人工智能系統中存在着偏見，但是有偏見的算法系統並不是一個新現象。隨着包括司法和健康等領域在內的各種組織都在採用人工智能技術，人們開始關注對基於人工智能的決策缺乏問責制和偏見。從人工智能研究人員和軟件工程師到產品領導者和消費者，各種各樣的利益相關者都參與到人工智能流水線中。在人工智能、數據集以及政策和權利領域的必要專業知識，可以共同揭示偏見，但是，這些利益相關者之間並不是統一可用的。因此，人工智能系統中的偏見會在不明顯的情況下複合。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"例如，機器學習開發人員，他們被要求: 對數據進行適當的預處理，從幾個可用的模型中選擇正確的模型，調整參數，調整模型體系結構以適應應用程序的需求。假設一個機器學習開發者被委託開發一個人工智能模型來預測哪些貸款會違約。由於沒有意識到訓練數據中的偏差，工程師可能會無意中只使用驗證的準確性來訓練模型。假設培訓數據中包含了太多違約的年輕人。在這種情況下，該模型很可能對年輕人在應用於測試數據時的違約行爲做出類似的預測。因此，機器學習的開發人員有必要了解可能潛入人工智能流水線的各種偏差以及導致的偏見。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/c9/c9e891dbdac09731496c7ffd2c5e2c84.jpeg","alt":"圖片.jpg","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在人工智能系統中定義、檢測、測量和減少偏見並不是一件容易的事情，而且是一個熱門的研究領域。各國政府、非營利組織和各行業都在做出許多努力，包括執行法規以解決與偏見有關的問題。認識和解決各種社會機構中的偏見，需要經過不斷的努力，以確保計算系統的設計，以解決這些問題。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這裏不對設計公平的人工智能算法提出建設性思考，而是在實踐方面，在數據創建，數據分析和評估的過程中，關注偏差與偏見的問題形成，，具體包括:","attrs":{}}]},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"人工智能流水線中的偏差分類。提供了各種類型偏差的結構組織，錨定在從數據創建和問題制定到數據準備與分析的各個階段。","attrs":{}}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"面向研究與實踐之間差距的建設性思路。分析在現實世界中實施研究的相關挑戰，並列出了填補這一空白的建議，希望可以幫助機器學習的開發者測試各種各樣的偏差。","attrs":{}}]}]}],"attrs":{}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"典型的人工智能流水線從數據創建階段開始: (1)收集數據; (2)對數據進行註釋或標記; (3)將數據準備或處理成其他管道可以使用的格式。讓我們分析在每個步驟中如何引入了不同類型的偏差。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/34/34ecf1ac07446f967fe8923b1ff74753.jpeg","alt":"圖片.jpg","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":1},"content":[{"type":"text","text":"數據集創建偏差","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在數據集的創建過程中，可能會出現特定類型的偏差。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"採樣偏差","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"通過選擇特定類型的實例而不是其他類型的數據集所產生的偏差稱爲採樣偏差。這是最常見的數據集偏差類型之一。例如，圖像數據集更喜歡街景或自然場景。人臉識別算法可能會得到更多淺膚色人臉的照片，從而導致識別深膚色人臉的偏差。因此，採樣偏差可能導致學習算法的泛化能力變差。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"測量偏差","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"測量偏差是由於人類測量中的誤差，或者由於人們在獲取數據時的某些固有習慣而引起的。例如，考慮圖像和視頻數據集的創建，其中的圖像或視頻可能反映了攝影師使用的技術。一些攝影師可能傾向於以類似的方式拍攝物體; 因此，數據集可能只包含特定角度的物體視圖。這種類型的測量偏差稱爲捕獲偏差。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"測量偏差的另一個來源可能是用於捕獲數據集的設備誤差。例如，用於捕捉圖像的相機可能存在缺陷，導致圖像質量差，從而導致有偏見的結果。這些類型的偏見又被廣泛地歸類爲設備偏見。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"當在創建數據集時使用代理而不是真實值時，可能會出現第三種測量偏差。例如，把醫生和用藥用來作爲醫療條件等的指標。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"標籤偏差","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"標籤偏差與標籤過程中的不一致性有關。不同的標註者有着不同的樣式和偏好，這些都反映在創建的標籤中。當不同的標註者爲同一類型的對象分配不同的標籤時，標籤偏見的一個常見例子就出現了。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"當評價者的主觀偏見影響標籤時，另一種類型的標籤偏見也會發生。例如，在詮釋文本中所體驗到的情感任務中，標註者的主觀偏好，如他們的文化、信仰和內省能力，可能會使標籤產生偏見。確認偏見，即人類傾向於搜索、解釋、關注和記憶信息以確認自己的先入之見，與這種類型的標籤偏見密切相關。因此，標籤可能是根據先前的信念而不是客觀的評估來分配的。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"第三種類型的標籤偏見可能產生於峯終效應。這是一種與記憶相關的認知偏見，人們在判斷一段經歷時，主要基於他們在經歷的頂峯(即最激烈的時刻)和結束時的感受，而不是基於這段經歷每一時刻的總和或平均值。例如，在分配標籤時，一些標準者可能更重視對話的最後一部分，而不是整個會話。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"否定集偏差","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"否定集偏差定義爲由於沒有足夠的代表“世界其他地方”的樣本而引入數據集的結果。數據集定義一個現象(例如，對象，場景，事件)不僅僅是根據它是什麼(正面的實例) ，還根據它不是什麼(負面的實例)。因此，分類器可能在檢測負實例方面表現不佳。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"問題定義產生的偏差","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"偏見還會根據問題的定義而產生。假設一家銀行想使用人工智能來預測客戶的信用可靠性。爲了做到這一點，必須以一種可以“預測或估計”的方式來定義信用可靠性這個問題，可以根據公司的需要來制定，比如說，最大化利潤率或最大化得到償還的貸款數量。然而，這些決定是出於各種商業原因，而不是公平或歧視。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"信用可靠性例子也可以被認爲是一種框架效應偏差。基於問題是如何表述的以及信息是如何呈現的，所得到的結果可能是不同的，甚至可能是有偏見的。因此，基於問題及其成功度量的定義方式，可能會產生偏差。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/c2/c218a421d9bd15f1762a13cd095d5255.jpeg","alt":"圖片.jpg","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":1},"content":[{"type":"text","text":"與算法/數據分析有關的偏差","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在算法或數據分析過程中可能會出現幾種類型的偏差。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"樣本選擇偏差","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"樣本選擇偏差是通過選擇個體、羣體或數據進行分析而引起的，這種方式使得樣本不能代表要分析的總體。特別地，樣本選擇偏差是在數據分析過程中由於對數據集中的某些變量(例如，特定的膚色、性別等)進行調節而產生的，這反過來又會產生虛假的相關性。例如，在分析母親身份對工資的影響時，如果僅限於已經就業的婦女，那麼由於條件作用在就業婦女身上，測量的效果就會有偏差。常見的樣本選擇偏差類型包括伯克森悖論和樣本截斷。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"混雜偏差","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在人工智能模型中，如果算法沒有考慮數據中的所有信息，或者沒有考慮特徵和目標輸出之間的關聯，從而學習了錯誤的關係，就會產生偏差。混雜偏差源於影響輸入和輸出的常見原因。一種特殊類型的混雜偏差是省略變量，它發生在一些相關的特徵沒有包含在分析中。這也與模型欠擬合問題有關。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"另一種類型的混雜偏見是代理變量。即使決策時不考慮敏感變量，分析中使用的某些其他變量也可以作爲這些敏感變量的“代理”。例如，郵政編碼可能表示民族，因爲某個民族的人可能主要居住在某個地區。這種偏見通常也被稱爲間接偏見或間接歧視。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"與設計有關的偏查","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"有時，由於算法的限制或系統的其他限制(如計算能力) ，也會出現偏差。在這個類別中一個值得注意的是算法偏差，它可以被定義爲僅由算法誘導或添加的偏差。依賴於隨機性來公平分配結果的軟件並不是真正的隨機，例如，通過將所選內容向列表末尾或開頭的選項傾斜，結果可能會有偏差。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"另一種與設計相關的偏差是排名偏差。例如，搜索引擎顯示每個屏幕三個結果，可以理解爲前三個結果的特權稍多於後三個。排名偏差也與表示偏差密切相關，這種偏差源於這樣一個事實，即你只能收到呈現給用戶的內容反饋。即使在那些已經顯示的內容中，收到用戶反饋的可能性也會受到該內容顯示位置的影響。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/60/6090808fe5283fc5055428a02da0e447.jpeg","alt":"圖片.jpg","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":1},"content":[{"type":"text","text":"與評價/驗證相關的偏差","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"有幾種類型的偏差源於人類評價者的固有偏差，以及在選擇這些評價者時的偏差。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"人類評估偏差","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"通常，人工評估者被用來驗證人工智能模型的性能。諸如確認偏差、峯終效應和先驗信念(如文化)等現象會在評估中產生偏差。人類評估者也會受到他們能回憶多少信息的限制，這可能會導致召回偏差。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"樣本處理偏差","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"例如，在推薦系統中，一些特定的觀衆(例如，那些說某種語言的人)可能會看到一則廣告，而另一些則不會。因此，觀察到的影響將不能代表對一般人羣的真正影響。在選擇性地對一些人羣進行某種處理的過程中引入的偏差稱爲樣本處理偏差。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"驗證和測試的數據偏差","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"一般而言，與數據集創建階段有關的偏差也可能出現在模型評估階段。此外，評估偏差可能來自於選擇不適當的基準/數據集進行測試。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"儘管在人工智能領域做了大量的研究工作來應對與偏見相關的挑戰，但是一些差距阻礙了進步。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":1},"content":[{"type":"text","text":"研究與實踐之間的差距","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"已經提出瞭解決數據集偏見問題的方法，新的數據集也在強調保持多樣性。例如，臉部多樣性數據集包括近100萬張從知識共享數據集中提取的人臉圖像，這些圖像是專門爲了實現膚色、臉部結構、年齡和性別之間的統計平等而組合起來的。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"“機器學習中的公平性”是一個活躍的研究領域。還有一些開放源碼工具，如 IBM 的 AI Fairness 3605，有助於檢測和減少不必要的算法偏差。儘管做出了這些努力，但仍然存在明顯的差距。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"爲了減少人工智能系統中潛在的偏見，已經提出了一些實踐指南。例如，建議使用具有詳細文檔的已發佈模型，並鼓勵透明度，需要創建特定於領域的教育資源、指標、流程和工具。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/a9/a998b475b9f5d191ba0dace7b86e4cbb.jpeg","alt":"圖片.jpg","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":1},"content":[{"type":"text","text":"對機器學習開發者的建議","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"雖然不可能消除所有的偏見來源，但是採取某些預防措施，可以減少一些偏見問題。以下建議可以幫助機器學習開發者識別潛在的偏見來源，並幫助避免不必要的偏見引入:","attrs":{}}]},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"納入特定領域的知識對於界定和發現偏見至關重要。理解數據集中各種特徵之間的結構依賴關係非常重要。通常，繪製一個結構圖來說明感興趣的各種特性及其相互依賴關係是有幫助的。這可以幫助我們找到偏見的來源。","attrs":{}}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"同樣重要的是，要根據應用程序瞭解哪些數據特徵被認爲是敏感的。例如，年齡可能是決定誰能得到貸款的一個敏感特徵，但不一定決定誰能得到醫療服務。此外，可能有一些代理特徵，雖然不被認爲是敏感特徵，但仍可能編碼敏感信息，從而使預測出現偏差。","attrs":{}}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"用於分析的數據集應儘可能代表真相。因此，在構建具有代表性的數據集時必須小心謹慎。","attrs":{}}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"必須明確適當的標準，以便爲數據作標註。規則的定義必須儘可能使標註者獲得一致的標籤。","attrs":{}}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"確定所有可能與目標特徵有關的特徵是重要的。省略與目標特性有依賴關係的變量會導致有偏差的估計。","attrs":{}}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"與輸入和輸出相關的特徵可能導致有偏差的評估。在這種情況下，重要的是通過適當的數據調節和選擇輸入的隨機化策略來消除這些偏差的來源。","attrs":{}}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"將數據分析限制在數據集的某些部分，可能會導致不必要的選擇偏差。因此，在選擇用於分析的數據子集時，必須注意不要引入樣本選擇偏差。","attrs":{}}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在驗證 a/b 測試等模型的性能時，必須注意防止引入樣本處理偏差。換言之，在測試模型的性能時，測試條件不應侷限於總體的某個子集。","attrs":{}}]}]}],"attrs":{}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/04/04545a5e401c0f0cb1c83ed563421f4b.jpeg","alt":"圖片.jpg","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":1},"content":[{"type":"text","text":"小結","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"從數據集的創建到問題的形成，從數據分析到結果的評估，人工智能流水線中可能出現各種偏差。一些經驗準則，可以幫助機器學習開發人員識別潛在的偏見來源，以及避免引入不必要的偏見。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"原載於公衆號「喔家ArchiSelf」","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":">>期待你的加入","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"百度開發者中心已開啓徵稿模式，歡迎開發者登錄","attrs":{}},{"type":"link","attrs":{"href":"https://developer.baidu.com/?from=010803","title":"","type":null},"content":[{"type":"text","text":"https://developer.baidu.com/?from=010803","attrs":{}}]},{"type":"text","text":"進行投稿，優質文章將獲得豐厚獎勵和推廣資源。","attrs":{}}]}]}