谷歌搜索引入多任务统一模型MUM，可更准确理解信息 ｜ 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":"美国时间 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开发者大会"}]},{"type":"text","text":"正式开幕。去年，该会议因疫情取消，今年重新恢复并采用全程线上的形式，对所有开发者免费开放。在刚刚结束的主题演讲中，谷歌发布了 TPU V4 人工智能芯片、"},{"type":"link","attrs":{"href":"https:\/\/www.infoq.cn\/article\/76gYqPA2YU0YXCDHFvIE","title":"xxx","type":null},"content":[{"type":"text","text":"自然语言平台 LaMDA "}]},{"type":"text","text":"以及一系列原有产品的更新升级。本文，我们将详细介绍谷歌搜索引入的多任务统一模型MUM。"}]}]},{"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":"每天都有很多人使用谷歌来处理需要多步骤的各种任务，而人们在处理类似的复杂任务时平均会发出 8 个查询。如今，搜索引擎还没有成熟到可以像专家一样回答问题。但随着“多任务统一模型”（Multitask Unified Model，MUM）的出现，谷歌正在帮助解决这类复杂需求。因此，未来只需要较少的搜索就可以完成任务。"}]},{"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":"与 "},{"type":"link","attrs":{"href":"https:\/\/blog.google\/products\/search\/search-language-understanding-bert\/","title":null,"type":null},"content":[{"type":"text","text":"BERT"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" 一样，MUM 同样基于 "},{"type":"link","attrs":{"href":"https:\/\/ai.googleblog.com\/2017\/08\/transformer-novel-neural-network.html","title":null,"type":null},"content":[{"type":"text","text":"Transformer 架构"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，但是它的功能要强大 1000 倍。MUM 不仅可以理解语言，而且可以生成语言。MUM 同时用 75 种不同的语言进行了多项任务的训练，使其比以前的模型更全面地理解信息和世界知识。此外，MUM 是多模态的，因此它能够理解文本和图像中的信息，将来，还可以扩展到视频和音频等更多模态。"}]},{"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":"以徒步富士山的问题为例：MUM 可以理解你在比较两座山，因此海拔高度和路径信息可能是相关的。它还可以理解，就远足而言，“准备工作”可能包括诸如健身训练以及寻找合适的装备。"}]},{"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":"因为 MUM 能够基于其对这个世界的深刻理解来展现自己的见解，所以它可以强调，尽管两座山的海拔高度大致相同，但秋季是富士山的雨季，你可能需要一件防水夹克。MUM 也能为更深层次的探索提供有用的副主题：比如顶级装备或最佳训练练习，并提供一些网络上有用的文章、视频和图片的链接。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/d9\/cf\/d9a8059879eb1ba4d4bea9137e35f6cf.gif","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"消除语言障碍"}]},{"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":"在获取信息时，语言可能是一个重要障碍。通过不同语言的知识迁移，MUM 有可能打破这些界限。它可以从那些不是用你的搜索语言写成的资料中学习，并且能帮助把这些相关信息发给你。"}]},{"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":"假设有一些关于富士山真正有用的信息是用日语写的；现在，如果你不用日语搜索，你很可能无法找到这些信息。然而，MUM 可以从不同语言的来源中迁移知识，并利用这些洞察力发现与你的首选语言最相关的结果。所以，在将来，当你搜索有关远足富士山的信息时，你可能会看到这样的结果：在何处能欣赏到富士山最美的风景、当地的温泉，以及受欢迎的纪念品商店……这些信息很容易用日语搜索就能找到。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/9e\/64\/9ea06b15a094bb4899c9b5d740a47164.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":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"理解不同类型的信息"}]},{"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":"多模态的 MUM 意味着它能够同时理解来自不同格式的信息，比如网页、图片等等。最终，你可能会拍一张登山靴的照片，然后问：“我能用它去爬富士山吗？”MUM 将会理解这张图片，并把它和你的问题联系在一起，让你知道你的靴子会很好用。之后，它会给你发一个博客网址，上面有推荐的装备列表。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/5c\/a7\/5cff98c01a53d22d31e91bcf703e8fa7.gif","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"带着负责的态度把高级人工智能运用到搜索中"}]},{"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":"无论何时，当我们使用人工智能来使世界上的信息更容易获取时，我们都要负责任地这样做。对于谷歌搜索的每一项改进，我们都会进行严格的评估，以确保我们能提供更加相关和有用的结果。那些遵循我们《"},{"type":"link","attrs":{"href":"https:\/\/static.googleusercontent.com\/media\/guidelines.raterhub.com\/en\/\/searchqualityevaluatorguidelines.pdf","title":null,"type":null},"content":[{"type":"text","text":"搜索质量评分准则"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"》（"},{"type":"text","marks":[{"type":"italic"},{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"Search Quality Rater Guidelines"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"）的人类评分者，帮助我们了解我们的结果如何帮助人们找到信息。"}]},{"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":"就像我们已经仔细测试了 BERT 从 2019 年开始推出的许多应用一样， MUM 也会经历同样的过程，将这些模型应用于搜索。具体地说，为了避免在我们的系统中引入偏见，我们将寻找可能显示机器学习中偏见的模式。同时，我们也会运用最新的"},{"type":"link","attrs":{"href":"https:\/\/arxiv.org\/abs\/2104.10350","title":null,"type":null},"content":[{"type":"text","text":"研究成果"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，比如如何减少 MUM 等训练系统的碳足迹，以确保搜寻工作尽可能高效。"}]},{"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":"今后数月甚至数年，我们将把 MUM 驱动的功能和改进带到我们的产品中。虽然我们仍处在 MUM 探索的初期，但这是一个重要的里程碑，将来谷歌能够理解人们自然地交流和解释信息的各种方式。"}]},{"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":"作者介绍："}]},{"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":"Pandu Nayak，谷歌研究员兼搜索部门副总裁。"}]},{"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":"原文链接："}]},{"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:\/\/blog.google\/products\/search\/introducing-mum\/"}]}]}