谷歌發佈自然語言平臺 LaMDA，新的基於AI的對話技術 | Google I/O 2021

{"type":"doc","content":[{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"美國時間5月18日，"},{"type":"link","attrs":{"href":"https:\/\/events.google.com\/io\/?lng=zh-CN","title":"xxx","type":null},"content":[{"type":"text","text":"Google I\/O 2021開發者大會"}]},{"type":"text","text":"正式開幕。去年，該會議因疫情取消，今年重新恢復並採用全程線上的形式，對所有開發者免費開放。在剛剛結束的主題演講中，谷歌發佈了TPU V4人工智能芯片、自然語言平臺 LaMDA以及一系列原有產品的更新升級。本文，我們將詳細介紹谷歌翻譯和自然語言平臺 LaMDA的主要特點。"}]}]},{"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":"color","attrs":{"color":"#494949","name":"user"}}],"text":"谷歌總是對語言情有獨鍾。早期，谷歌就着手建立了翻譯網絡。近年來，谷歌開始利用機器學習技術"},{"type":"link","attrs":{"href":"https:\/\/blog.google\/products\/search\/search-language-understanding-bert\/","title":null,"type":null},"content":[{"type":"text","text":"更好地理解搜索查詢的意圖"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"。隨着時間的推移，谷歌在這些領域取得的進展使得用書面和口頭語言組織和獲取大量信息變得更加容易。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/1e\/98\/1e54388498d2d9651b366e30ac2be698.gif","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","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"但是，技術總有改進的空間。語言具有顯著的細微差別和適應性。它可以是字面的，也可以是語音的；可以是華麗的，也可以是樸素的；可以是創意性的，也可以是信息性的。這種多功能性使得語言成爲人類最偉大的工具之一，也是計算機科學最難解決的問題之一。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"作爲最新的研究突破， LaMDA 爲這個難題中最吸引人的部分增加了一些內容：對話。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/5f\/73\/5f0f5649b1d7f7a8567f5e34d0791e73.gif","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","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"儘管對話往往圍繞特定主題進行，但是對話的開放性意味着對話可以從一個地方開始，到另一個完全不同的地方結束。和朋友聊到一個電視節目，可能會演變成一場關於這個節目拍攝國家的討論，然後轉而討論這個國家最好的地方美食。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"這種特性很快就會讓現代對話智能體（通常被稱爲聊天機器人）陷入困境，因爲它們經常遵循狹窄的、預定路徑。但是 LaMDA（Language Model for Dialogue Applications)的縮寫，意爲“對話應用語言模型”）能夠以一種自由流動的方式討論無止境的主題，我們認爲，這一能力可以使與技術的交互更加自然，並提供一種全新類別的應用程序。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"通向 LaMDA 的路道阻且長"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"LaMDA 的對話技能已經醞釀多年。與包括 BERT 和 GPT-3 在內的許多最新語言模型一樣，它建立在 "},{"type":"link","attrs":{"href":"https:\/\/ai.googleblog.com\/2017\/08\/transformer-novel-neural-network.html","title":null,"type":null},"content":[{"type":"text","text":"Transformer"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" 上，這是由谷歌研究院發明並於 2017 年開源的一個神經網絡架構。由這個架構生成的模型可以訓練閱讀許多單詞（例如，一個句子或段落），注意這些單詞之間的關係，然後預測它認爲接下來會出現什麼單詞。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"但與其他大多數語言模型不同的是，LaMDA 接受的是對話訓練。在訓練過程中，它發現一些區別於其他語言形式的開放式對話的細微差異。合理性是其中的一個細微差異。基本上是這樣：對特定對話環境的反應是否具有意義？舉例來說，如果有人說："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"“I just started taking guitar lessons.”"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"你也許希望別人會這樣回答："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"“How exciting! My mom has a vintage Martin that she loves to play.”"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"從最初的陳述來看，這種迴應是有意義的。但合理並非良好反應的唯一因素。畢竟，“that's nice” 這句話幾乎是對任何陳述句的合理迴應，正如 “I don't know” 是對大多數問題的合理迴應一樣。令人滿意的答覆通常也是具體的，與對話的上下文密切相關。在上面的例子中，迴應是合理且具體的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"LaMDA 建立在谷歌 "},{"type":"link","attrs":{"href":"https:\/\/ai.googleblog.com\/2020\/01\/towards-conversational-agent-that-can.html","title":null,"type":null},"content":[{"type":"text","text":"2020 年發表"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"的早期研究之上，該研究表明，基於 Transformer 的語言模型經過對話訓練，可以學會談論幾乎任何事情。此後，我們還發現，一旦經過訓練，LaMDA 可以進行微調，從而大幅提高其反應的合理性和特異性。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"目前還處於早期發展階段，我們希望不久能有更多分享，但是合理性和特異性並非我們在 LaMDA 這樣的模型中所尋求的唯一特性。通過評估迴應是有洞察力的、意想不到的還是機智的，我們也在探索像“趣味性”這樣的維度。谷歌也非常關注事實性（即 LaMDA 是否堅持事實，這是語言模型經常遇到的問題），並且正在研究如何確保 LaMDA 的反應不僅有說服力，而且正確。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"但對於我們的技術，我們會問自己一個最重要的問題，那就是它們是否符合"},{"type":"link","attrs":{"href":"https:\/\/www.blog.google\/technology\/ai\/ai-principles\/","title":null,"type":null},"content":[{"type":"text","text":"我們的人工智能原則"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"。文字可能是人類最偉大的工具之一，但是和其他一切工具一樣，它也會被濫用。這種濫用可以通過受過語言訓練的模型來傳播，例如，把偏見內化，反映出仇恨的言論，或者複製誤導信息。儘管模型所訓練的語言已經被仔細地審查過了，但模型本身仍有被濫用的危險。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"在創建 LaMDA 這樣的技術時，我們的首要任務是努力確保將這種風險降至最低。由於我們多年來致力於研究和發展這些技術，所以我們對機器學習模型所涉及到的問題非常熟悉。正因爲如此，我們建立並開源資源和數據，讓研究人員可以用來分析模型和訓練模型；我們在 LaMDA 開發的每一個步驟都仔細檢查過；我們承諾在更多產品中增加對話能力，所以我們會繼續這樣做。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"作者介紹："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"Eli Collins，谷歌產品管理副總裁。Zoubin Ghahramani，谷歌高級研究總監。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"原文鏈接："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"https:\/\/www.blog.google\/technology\/ai\/lamda"}]}]}