爲什麼神經網絡不適合理解自然語言 ？

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"過去十年中，人工智能行業的一大趨勢就是通過創建更大的深度學習模型來解決問題。這種趨勢在自然語言處理領域最爲明顯，這也是人工智能最具挑戰性的領域之一。"}]},{"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":"近年來，研究人員已經證明，向神經網絡添加參數可以提高它們在語言任務上的表現。然而，語言理解的根本問題——單詞和句子下隱藏的，名爲含義的冰山——仍有待解決。"}]},{"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":"倫斯勒理工學院兩位科學家的著作《"},{"type":"link","attrs":{"href":"https:\/\/mitpress.mit.edu\/books\/linguistics-age-ai","title":"","type":null},"content":[{"type":"text","text":"人工智能時代的語言學"}]},{"type":"text","text":"》討論了當前各種自然語言理解（NLU）方法的缺點，並探索了一些開發先進智能代理的未來途徑——這些智能代理可以與人類自然交互，而不會讓交流陷入困境或犯愚蠢的錯誤."}]},{"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":"《人工智能時代的語言學》的作者Marjorie McShane和Sergei Nirenburg認爲，人工智能系統不能止步於對單詞的操縱。在他們的書中，他們證明了NLU系統可以理解世界，向人類解釋它們獲得的知識，並在它們探索世界時不斷學習。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"基於知識的系統與知識精益的系統"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/resource\/image\/fd\/da\/fdc4851c15057d2093af47c0b9157dda.jpg","alt":null,"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":"考慮這句話，“I made her duck.”這句話的主題是扔了一塊石頭讓對方彎下腰，還是他給她煮了鴨肉？"}]},{"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":"現在再考慮這句話：“Elaine poked the kid with the stick.”Elaine是用棍子戳了那個孩子，還是用她的手指戳了碰巧拿着棍子的孩子？"}]},{"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":"語言充滿了歧義。我們人類使用語言的上下文來解決這些歧義。我們使用來自說話者的語氣、先前的單詞和句子、對話的一般性設置以及關於世界的基本知識等線索來建立上下文。當我們的直覺和知識未能解決歧義時，我們會提出問題。對我們來說，確定上下文的過程是很容易的。但要以可計算的方式定義這種過程，那就是說起來容易做起來難了。"}]},{"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":"通常有兩種方法可以解決這個問題。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/resource\/image\/df\/35\/df532faf681b7d4827f1ba1d0fa70135.jpg","alt":null,"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":"center","origin":null},"content":[{"type":"text","text":"人工智能時代的語言學——Marjorie McShane和Sergei Nirenburg"}]},{"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":"在之前幾十年的人工智能研究中，科學家使用基於知識的系統來定義句子中每個單詞的作用，並以此提取句子的上下文和含義。基於知識的系統依賴於大量關於語言、情境和世界的特徵。這些信息可以來自不同的來源，並且必須以不同的方式計算。"}]},{"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":"基於知識的系統提供了可靠且可解釋的語言分析結果。但它們最後失寵了，因爲它們需要太多的人力來設計特徵、創建詞彙結構和本體，和開發將所有這些部分結合在一起的軟件系統。研究人員將知識工程中的人工環節視爲一種瓶頸，並尋求其他方法來應對語言處理過程。"}]},{"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":"“人們普遍認爲，克服這種所謂知識瓶頸的任何嘗試都是徒勞的；而這種看法也深刻地影響了通用人工智能，尤其是NLP[自然語言處理]的發展道路，使該領域遠離了理性主義和基於知識的方法，並導致了NLP中經驗主義、知識精益、研究和開發範式的出現，”McShane和Nirenburg在《人工智能時代的語言學》中寫道。"}]},{"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":"近幾十年來，機器學習算法一直是NLP和NLU的核心。機器學習模型是一種知識精益（knowledge-lean）系統，它試圖通過統計關係來處理上下文問題。在訓練期間，機器學習模型處理大量文本，並根據單詞彼此之間的位置關係調整其參數。在這些模型中，上下文是由單詞序列之間的統計關係，而不是單詞背後的含義來決定的。自然，數據集越大、示例越多樣化，這些數值參數就越能捕捉單詞彼此之間的各種位置組合。"}]},{"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":"知識精益系統之所以能流行，主要歸功於可用來訓練機器學習系統的大量計算資源和大型數據集。藉助維基百科等公共數據庫，科學家們能收集到龐大的數據集，並針對翻譯、文本生成和問答等各種任務訓練他們的機器學習模型。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"機器學習不計算含義"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如今，我們的深度學習模型可以"},{"type":"link","attrs":{"href":"https:\/\/bdtechtalks.com\/2020\/09\/14\/guardian-gpt-3-article-ai-fake-news\/","title":"","type":null},"content":[{"type":"text","text":"生成文章篇幅的文本序列"}]},{"type":"text","text":"、"},{"type":"link","attrs":{"href":"https:\/\/bdtechtalks.com\/2019\/09\/09\/aristo-ai-science-exam\/","title":"","type":null},"content":[{"type":"text","text":"回答科學考試問題"}]},{"type":"text","text":"、"},{"type":"link","attrs":{"href":"https:\/\/bdtechtalks.com\/2021\/07\/05\/openai-github-gpt-3-copilot\/","title":"","type":null},"content":[{"type":"text","text":"編寫軟件源代碼"}]},{"type":"text","text":"以及回答基本的客戶服務諮詢問題。由於深度學習架構的種種改進（LSTM、transformer），更重要的是由於神經網絡每年都在變大，這些領域中的大多數都取得了進展。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/resource\/image\/c8\/ef\/c8e27dc409f5313abbd1ed45dc33edef.jpg","alt":null,"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":"center","origin":null},"content":[{"type":"text","text":"近年來，深度學習語言模型的規模不斷擴大（以對數尺度製圖）"}]},{"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":"但是，雖然更大的深度神經網絡可以爲許多任務類型提供增量改善，但它們並沒有從宏觀層面上解決自然語言理解的一般性問題。這就是爲什麼各種實驗都表明，即使是最複雜的語言模型也無法解決關於世界是如何運作的一些"},{"type":"link","attrs":{"href":"https:\/\/www.technologyreview.com\/2020\/08\/22\/1007539\/gpt3-openai-language-generator-artificial-intelligence-ai-opinion\/","title":"","type":null},"content":[{"type":"text","text":"簡單問題"}]},{"type":"text","text":"。"}]},{"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":"在他們的書中，McShane和Nirenburg將當前人工智能系統所解決的問題描述爲“唾手可得的果實”。一些科學家認爲，神經網絡繼續擴展下去，終有一天會解決機器學習所面臨的問題。但McShane和Nirenburg認爲我們需要解決一些更本質的問題。"}]},{"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":"“這樣的系統並不像人類那樣思考：它們不知道自己在做什麼以及爲什麼這樣做，它們解決問題的方法與人類不同，而且它們不依賴於世界、語言或代理的模型，”他們寫道。“相反，它們在很大程度上依賴於將通用機器學習算法應用於更大數據集的路徑，並得到了現代計算機驚人的速度和存儲容量的支持。”"}]},{"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TechTalks上發表的評論中，認知科學家和計算語言學家McShane表示，機器學習必須克服幾個障礙，其中首當其衝的是"},{"type":"link","attrs":{"href":"https:\/\/bdtechtalks.com\/2020\/07\/13\/ai-barrier-meaning-understanding\/","title":"","type":null},"content":[{"type":"text","text":"含義的缺失"}]},{"type":"text","text":"。"}]},{"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":"“統計\/機器學習（S-ML）方法並不會去計算含義，”McShane說。“相反，從業者一路向前，就好像只憑單詞就足以代表句子的含義一樣，而事實並非如此。實際上，當涉及到句子的完整上下文含義時，句子中的單詞只是冰山一角。將詞語與含義混淆的這種人工智能方法，就像一艘駛向冰山的巨輪一樣令人擔憂。”"}]},{"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":"在大多數情況下，機器學習系統通過縮小任務範圍或擴大訓練數據集來回避處理單詞含義的問題。但是，即使一個大型神經網絡設法在相當長的一段文本中保持了連貫性，但在背後，它也仍然無法理解它所生成的那些單詞的含義。"}]},{"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":"“當然，人們可以構建看起來表現得很聰明的系統（例如GPT-3），只不過這些系統真的不知道到底發生了什麼事情，”McShane說。"}]},{"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":"一旦你問一系列簡單但互相關聯的問題，所有基於深度學習的語言模型就會開始崩潰，因爲它們的參數無法捕捉日常生活中潛藏的無限複雜性。在這個問題上投入更多數據並不能將知識顯式集成到語言模型中。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"語言賦能的智能代理（LEIA）"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/resource\/image\/21\/60\/218abb1b1d15c0d3a8d4d40d332a8960.jpg","alt":null,"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":"center","origin":null},"content":[{"type":"text","text":"Marjorie McShane和Sergei Nirenburg，《人工智能時代的語言學》的作者"}]},{"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":"在他們的書中，McShane和Nirenburg提出了一種解決自然語言理解過程中“知識瓶頸”的方法，這種方法無需求助於需要大量數據的純機器學習手段。"}]},{"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":"《人工智能時代的語言學》的核心是稱爲“語言賦能的智能代理（LEIA）”的概念，其具有三個關鍵特徵："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"numberedlist","attrs":{"start":1,"normalizeStart":1},"content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":1,"align":null,"origin":null},"content":[{"type":"text","text":"LEIA可以理解語言的上下文相關含義，並從單詞和句子的歧義中找到合適的理解。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":2,"align":null,"origin":null},"content":[{"type":"text","text":"LEIA可以向它們的人類合作者解釋它們的想法、行動和決策。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":3,"align":null,"origin":null},"content":[{"type":"text","text":"與人類一樣，LEIA可以在與人類、其他代理和世界互動時進行終身學習。終身學習（Lifelong learning）減少了爲擴展智能代理的知識庫而持續投入人力的需求。"}]}]}]},{"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":"LEIA通過六個階段來處理自然語言，這些階段從確定單詞在句子中的作用到語義分析，最後是情境推理。這些階段讓LEIA可以解決單詞和短語的不同含義之間的衝突，並將句子整合到代理正在處理的更廣泛的上下文中。"}]},{"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":"LEIA爲它們對語言表達的各種解釋分配置信度，並且知道它們的技能和知識何時不足以解決歧義。在這種情況下，它們與人類同行（或它們環境中的智能代理和其他可用資源）互動以解決歧義。這些互動反過來又讓它們能夠學習新事物並擴展它們的知識。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/resource\/image\/be\/b4\/bed279e56892aaa048393f08yy0cb3b4.jpg","alt":null,"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":"center","origin":null},"content":[{"type":"text","text":"LEIA分幾個階段處理語言輸入"}]},{"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":"LEIA將句子轉換爲文本含義表示（TMR），這是對句子中每個單詞的可解釋和可操作的定義。LEIA根據它們的上下文和目標來確定需要跟進哪些語言輸入。例如，如果一個維修機器人與幾位人類技術人員共用一個機器維修車間，並且人類在討論昨天的體育比賽結果，那麼人工智能應該能夠分辨出哪些對話與其工作相關（機器維修），哪些是它可以忽略的（運動）。"}]},{"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":"LEIA傾向於使用基於知識的系統，但它們也在流程中集成了機器學習模型，尤其是在語言處理一開始的句子解析階段。"}]},{"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":"“我們很樂意集成更多S-ML引擎，只要它們能夠提供各種類型的高質量啓發式證據（但是，當我們合併黑盒S-ML結果時，代理的置信度估計和可解釋性都會受到影響），”McShane說。“我們也期待結合S-ML方法來執行一些面向大數據的任務，例如選擇示例來輔助閱讀學習過程。”"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"語言理解需要人腦複製品嗎？"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/resource\/image\/20\/d9\/2074af71b108795d65494d6fe91cf9d9.jpg","alt":null,"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":"LEIA的主要特徵之一是知識庫、推理模塊和感官輸入的集成。目前，計算機視覺和自然語言處理等領域之間幾乎沒有重疊。"}]},{"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":"正如McShane和Nirenburg在他們的書中指出的那樣，“語言理解不能與整體的代理認知過程區分開來，因爲支持語言理解的啓發式方法也要運用其他感知模式（例如視覺）生成的結果，來推理說話者的計劃和目標，並推理需要花費多少資源來理解困難的輸入。”"}]},{"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":"在現實世界中，人類利用他們豐富的感官體驗來填補語言表達的空白（例如，當有人對你說“看那邊？”時，他們假設你可以看到他們的手指指向的地方）。人類進一步開發了理解彼此思維的模型，並使用這些模型做出假設並忽略語言中的細節。我們希望任何以我們自己的語言與我們交互的智能代理都具有類似的能力。"}]},{"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":"“我們完全理解爲什麼現在孤立方法成了常態：每種問題解釋起來都很困難，每個問題的實質都需要單獨分析，”McShane說。“然而，如果沒有集成，所有問題的實質層面都無法解決，因此重要的是要抵制（a）假設模塊化必然會導致簡化，以及（b）無限期地推遲集成的想法。”"}]},{"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":"同時，實現類似人類的行爲並不需要LEIA成爲人類大腦的複製品。“我們同意"},{"type":"link","attrs":{"href":"https:\/\/philosophynow.org\/issues\/88\/Aping_Mankind_Neuromania_Darwinitis_and_the_Misrepresentation_of_Humanity_by_Raymond_Tallis","title":"","type":null},"content":[{"type":"text","text":"Raymond Tallis"}]},{"type":"text","text":"（和其他人）的觀點，即他所謂的神經躁狂症——渴望解釋作爲一個生物實體的大腦可以告訴我們哪些關於認知和意識的內容——導致了許多無法真正解釋的可疑主張和解釋，”McShane說。“至少在當前的發展階段，神經科學無法爲我們的認知建模類型和目標提供任何內容（句法或結構）支持。”"}]},{"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《人工智能時代的語言學》中，McShane和Nirenburg認爲複製大腦不符合AI的可解釋性目標。“運行在人類代理團隊中的[代理]，需要在一定程度上了解輸入，以確定它們應該追求哪些目標、計劃和行動，來作爲NLU的輸出結果，”他們寫道。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"一個長期目標"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/6c\/c1\/6cb6a7a81ef47d1d20f4f4e7424c32c1.jpg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"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":"《人工智能時代的語言學》中討論的許多主題仍處於概念層面，離實現還有很長的距離。作者提供了NLU的每個階段應該如何運作的藍圖，儘管實際的系統尚不存在。"}]},{"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":"但McShane對LEIA的發展持樂觀態度。“從概念和方法來說，工作進展都是非常順利的。主要障礙是在當前的行業氛圍下缺乏資源來分配給基於知識的方法，”她說。"}]},{"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":"McShane認爲，在批評基於知識的系統時，焦點都集中在知識瓶頸上，但其實這種批評在幾個方面都有誤導性："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"numberedlist","attrs":{"start":1,"normalizeStart":1},"content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":1,"align":null,"origin":null},"content":[{"type":"text","text":"實際上並不存在所謂的瓶頸，只要向前邁步就對了。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":2,"align":null,"origin":null},"content":[{"type":"text","text":"相關工作在很大程度上可以自動執行，可以讓代理通過自己的操作學習語言、瞭解世界，並由人類獲得的高質量核心詞典和本體引導代理。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":3,"align":null,"origin":null},"content":[{"type":"text","text":"儘管McShane和Nirenburg認爲AI代理可以自動學習多種知識——尤其是當引導代理的知識庫變得更大時——但最有效的知識獲取流程定然需要人工參與，這種參與可能是爲了質量控制或者處理困難用例等目的。"}]}]}]},{"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":"她說：“我們準備發起大規模的工作計劃來推動LEIA的採用，這將使涉及語言交流的各種應用程序更像人類。”"}]},{"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":"在他們的著作中，McShane和Nirenburg也承認我們需要做很多工作，且LEIA的發展是一項“持續的、長期的、範圍廣泛的工作計劃”。"}]},{"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":"“要做的工作的深度和廣度與目標的崇高程度是相稱的，這個目標就是：讓機器能夠像人類一樣熟練地使用語言，”他們在《人工智能時代的語言學》中寫道。"}]},{"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","marks":[{"type":"strong"}],"text":"原文鏈接："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/bdtechtalks.com\/2021\/07\/12\/linguistics-for-the-age-of-ai\/","title":"","type":null},"content":[{"type":"text","text":"https:\/\/bdtechtalks.com\/2021\/07\/12\/linguistics-for-the-age-of-ai\/"}]}]}]}