人物访谈 | Janelle Shane博士访谈
点击上方蓝字关注我们
编者按
《理论语言学五道口站》（2023年第32期，总第296期）人物专栏与大家分享Lauren Gawn博士和网络语言学家Gretchen McCulloch对Janelle Shane博士的访谈。Lauren Gawn，澳大利亚乐卓博大学语言与语言学系高级讲师。Gretchen McCulloch，加拿大网络语言学家。Janelle Shane博士，美国Meadowlark Optics公司高级科学家，人工智能研究员。
本期访谈节选自播客节目Lingthusiasm对Janelle Shane博士进行的专访。在访谈中，Janelle Shane博士根据自己的人工智能研究经验回答了人工智能语言学习的相关问题。访谈内容转自网站：https://lingthusiasm.com，由本站成员黄静雯、安安翻译。
采访人物简介
Janelle Shane 博士
Janelle Shane，光学研究科学家及人工智能研究员，美国Meadowlark Optics公司高级科学家，美国加利福尼亚大学圣迭戈分校电气工程博士。她是AI Weirdness网站的创始人，其中记录了很多机器学习相关的算法。著有You Look Like A Thing And I Love You: How AI Works And Why It's Making The World A Weirder Place一书，已于2019年11月5日在Hachette Audio出版。
Brief Introduction of Interviewee
Janelle Shane is an optics research scientist and artificial intelligence researcher, working as a senior scientist at Meadowlark Optics Inc. She is an electrical engineering PhD at University of California, San Diego. She is the founder of AI Weirdness, where she documents various machine learning algorithms. She is the author of You Look Like A Thing And I Love You: How AI Works And Why It's Making The World A Weirder Place, which was published by Hachette Audio on Nov 5th, 2019.
采访者简介
Lauren Gawne 博士
Lauren Gawne，澳大利亚乐卓博大学语言与语言学系高级讲师，澳大利亚墨尔本大学语言学博士。她的研究兴趣为记录和分析人们说话和做手势的方式，目前的研究重点是手势使用上的跨文化差异。她与Gretchen McCulloch共同主持播客Lingthusiasm，并运营语言学网站Superlinguo。
Gretchen McCulloch
Gretchen McCulloch，加拿大网络语言学家，纽约时报畅销书Because Internet: Understanding the New Rules of Language的作者。她的研究课题主要是对线上交流（如互联网表情包、绘文字、即时通信等）进行语言学分析。她是Wired月刊和女性杂志网站The Toast的常驻语言学家以及Lingthusiasm播客的联合创立人。
Brief Introduction of Interviewers
Lauren Gawne is a Senior Lecturer in the Department of Languages and Linguistics at La Trobe University. Her linguistics PhD is from the University of Melbourne. She is interested in documenting and analyzing how people speak and gesture. Her current research focus is the cross-cultural variation in gesture use. She co-hosts the podcast Lingthusiasm with Gretchen McCulloch and run the linguistics website Superlinguo.
Gretchen McCulloch is a Canadian internet linguist and author of the New York Times bestselling Because Internet: Understanding the New Rules of Language. She offers linguistic analysis of online communication such as internet memes, emoji and instant messaging. She’s been the Resident Linguist at Wired and The Toast and is the co-creator of Lingthusiasm.
访谈内容
01.
Lauren Gawne博士：在您的研究中，您以特定的语体或术语集为对象收集大量语料，并将这些数据输入到人工智能中，随后它便会产出各种异想天开的有趣结果。可以说，人工智能生成结果的“灵感”都是来源于训练输入的数据。那么首先请您先来给我们讲讲人工智能为冰淇淋口味起的那些名字吧。
Janelle Shane博士：德克萨斯州奥斯汀市Kealing中学的一些学生参加了编程课，并决定要利用人工智能生成全新的冰淇淋口味名称。我得知了这一想法后，提示他们程序需要输入现有的冰淇淋口味名称以用于模仿，只有这样它才能知道冰淇淋口味的名称究竟是什么。因此，每名同学都从各个网站搜集并整理出了一份现有冰淇淋口味名称的丰富数据，包括多达1600种不同的口味。利用如此庞大的语料数据，他们成功得到了一些非常有趣的口味名称。
02.
Gretchen McCulloch：我有一篇博文专门谈到了这些中学生得到的冰淇淋口味名称，有一些起得非常好，如“周日来临（It’s Sunday）”、“樱桃诗人（Cherry Poet）”、“香脆芝士蛋糕（Brittle Cheesecake）”和“快乐蜂蜜香草（Honey Vanilla Happy）”。这样的冰淇淋口味似乎是合乎常理的。但同时也存在一些奇怪的口味，如“巧克力手指（Chocolate Finger）”、“焦糖书籍（Caramel Book）”以及更为奇怪的名称，如“化脓的坚果 （Nuts with Mattery）”、“香脆棕色 （Brown Crunch）”以及“曲奇与绿色 （Cookies and Green）”。
Lauren Gawne博士：这些名称的奇怪之处在于它们的语义。我们可以看到，尽管这些奇怪的结果仍然是由英文单词组成，但却仅仅只是看起来像英文单词而已。值得一提的是，人工智能根本不知道冰淇淋是什么，它只是利用已有的口味清单来判断什么样的词语组合看起来像是一个冰淇淋的口味。
Janelle Shane博士：是的，人工智能的发展还停留在非常基础的阶段。它们只能回答一些比较简单的问题，比如，什么字母倾向于出现在其它字母后面? 有哪些字母组合是常见的？又有哪些组合是不会出现的？ 因此，在训练过程中，它会在一些试错后学会如何正确地拼写像“巧克力”这样的常见单词，但这并不意味着它懂得“巧克力”这一概念。人工智能所要做的，或者说它获取进步的唯一方式，便是尝试去预测下一个应使用的字母或字母组合。之后，它会基于之前从未见过的实际文本来检验预测的准确性，并根据这一结果来决定是否需要对自己的内部结构进行调整以实现更为准确的产出。在这一过程中，人工智能在不断地试错，并反复对自己的猜测进行验证。
03.
Lauren Gawne博士：原来如此。它在进行拼写的时候总是会面临多种可能的拼写方案，例如“CH”和“HC”；而只有当它以“CH”对冰淇淋口味名单进行学习的时候，才会得到如“巧克力（chocolate）”、“脆片（chip）”和“樱桃（cherry）”这样的正向反馈。
Gretchen McCulloch：那么，我们是不是也可以认为人工智能并不了解世界的真实情况呢？因为我们可以看到，生成结果中也存在像“花生酱粘液（Peanut Butter Slime）”这样的名称；虽然都是英语单词，但对于冰淇淋的口味来说，这些单词的组合则是不可接受的。
Janelle Shane博士：是的。我们也需要明白的是，当前人工智能所拥有的计算能力远远不及人类所具备的感知能力水平。仅就基本的计算能力而言，人工智能的神经网络也只达到了蚯蚓的水平。
04.
Lauren Gawne博士：在您的工作中，除冰淇淋口味名称之外，您也曾试验生成死亡金属风格的音乐名称、万圣节装扮名称和颜色名；如同冰淇淋名称，这些名称的长度约为3-4个单词。目前您有用于生成几个词的人工智能，也有用于生成搭讪台词的人工智能。相较而言，搭讪台词则更接近句子或者句子的组合。当您试图生成更长的句子时，人工智能所面对的困难是否会有显著增加呢？
Janelle Shane博士：是的。在我的研究工作中，其中一个人工智能的记忆储存空间极为有限。在工作过程中，它的记忆能力仅限于记住几个单词，因此无法用于短语或句子的生成。搭讪台词的生成任务显然不是这一人工智能所能胜任的。同时，生成搭讪台词任务的挑战性也在于，人工智能不仅需要生成合法的句子，并且需要处理好句子中的双关语和暗示等内容。这些内容的处理都要求人工智能掌握大量的背景知识，而目前它们却仍不具备这样的能力。
05.
Gretchen McCulloch：您书中给出的另一个例子是关于食谱的生成。人工智能明白食谱的内容中应包含所需的食材，以及各加工步骤的具体操作信息，但它最终生成的食谱中却并不一定包含之前提到的食材，这是因为它并不记得自己之前所罗列的内容。
Janelle Shane博士：是的。我们会得到一个看上去像食谱但实际并不是食谱的产出结果。人工智能不是遗忘了食材的名称，就是在单纯复制人类所写的内容。因此人工智能在生成过程中只是在不断地猜测人类可能会说什么。
06.
Gretchen McCulloch：我想这也能解释您所提出的著名的“长颈鹿问题”。
Janelle Shane博士：“长颈鹿问题”出现在一个名为Visual Chatbot的人工智能程序中，这一人工智能是为了回答有关图像的问题而设计的。在这一研究中，训练数据对人工智能生成结果的影响也是至关重要的。对于它来说，设计者需要在训练数据中避免启动效应的问题。日常生活中，人们倾向于提出能获得肯定答复的问题。在该聊天机器人的早期版本中，设计者发现它仅仅是对所有的是否问句一律答“是”就获得了高达百分之八十的正确率。设计者提出的解决方案则为不向提问者展示具体图像信息，这就使得机器人在回答一个指定问题时回答“是”或“否”的概率均为百分之五十。但在这之后，设计者仍面临一个难题，即“长颈鹿问题”。设计者发现，在对机器人提问“你能在图像中看到几只长颈鹿？”时，机器人总是会给出一个非零答案，这便是它模仿人类答案的结果。对于人类来说，当我们不知道图中有长颈鹿时，我们是不会提这个问题的。
07.
Gretchen McCulloch：是的。我们倾向于先询问是否存在长颈鹿；而只有在得到肯定回答后，我们才会继续询问长颈鹿的数量。
Janelle Shane博士：非常正确。当我们向聊天机器人提出是否有长颈鹿的疑问时，它常常会回答“否”。但如果这时再对数量进行追问的话，却会得到“五”这样的确切数字。
08.
Lauren Gawne博士：您开发的命名程序也用于精酿啤酒品牌的命名。而人工智能所生成的结果最后也真的被一家啤酒公司所采用。您承接这一项目是想测试创意性命名的效果吗？还是说您有什么别的考虑呢？
Janelle Shane博士：现实生活中，曾有啤酒公司因为取名过于相似而将彼此告上法庭。因此，我们需要证明啤酒的命名仍未穷尽所有的可能性。而我们所做的就是与人工智能展开合作：我们对人工智能给出的名字进行整理、挑选，以便找到那些既遵循我们所给出的命名规则又具有一定发散性创意的名字。
09.
Lauren Gawne博士：目前为止，我们谈到了研究中为保证人工智能产出质量而进行的结果筛选工作；我们也谈及了为避免类如“长颈鹿问题”这一现象而适当选择输入数据类型的重要性。但我想，就创造完善的人工智能而言，挑战不仅限于对输入数据语言的选择，更在于对于输入内容主题的选择。您认为当前研究工作面对的最大挑战是什么？
Janelle Shane博士：我想最大的挑战在于当人工智能对输入语料进行学习时，由于它们自身的计算能力有限且缺乏外部语境知识，因此它们会对特定的人类活动进行不恰当的模仿，这其中就包括它们对种族或性别歧视言论的模仿。人工智能无法将其判断为错误行为，也无法理解其中的涵义。
10.
Lauren Gawne博士：在过去的几年中，人工智能对于长篇文本的处理能力似乎有所提升。请问确实如此吗？
Janelle Shane博士：是的。在2016年时，我所进行的项目仅限于生成单词、短语、油漆颜料名称、冰淇淋口味名称等内容长度较短的结果。那时，句子的生成相当困难，更遑论具备连续性的多个句子。但就在去年，接受过互联网数百万网页内容训练的大型人工智能已经诞生，它们处理长篇文本的能力也有了显著的提升。大部分情况下，即便这些人工智能仍然无法理解产出的内容，它们也可以生成以真实词语为基础且合乎语法的句子。因此，我认为人工智能的研究已经实现了长足的进步。
English Version
01.
Dr. Lauren Gawne: Janelle, in your work, you take large data sets of particular sets of terms or particular language genres, and then you feed them into an artificial intelligence, and then it spits out these delightfully whimsical outputs. It takes inspiration from the data set that it’s given. Let’s start with ice cream names generated by AI.
Dr. Janelle Shane: There’s a school in Austin, Texas, called Kealing Middle School where there is a group of students in the coding classes who decided that they wanted to generate ice cream flavors. I found out about their idea, and I said, “I need examples of existing ice cream flavors” because the A.I. has to have something to imitate. It doesn’t know about ice cream flavors unless I have some to tell it about. Then, each of them went and collected a few from this site or that site. They put together this amazing data set of existing ice cream flavors, about 1600 different ones. With the data set that big, they started generating pretty amusing flavors.
02.
Gretchen McCulloch: I’ve got the blogpost up about the ice cream flavors from the middle school students, and some of them are really good, like “It’s Sunday” and “Cherry Poet” and “Brittle Cheesecake” and “Honey Vanilla Happy.” These seem like kind of reasonable ice cream flavors. But there are also some weirder flavors from this data set like, “Chocolate Finger” and “Caramel Book”. Then, there’s this even weirder category, “Nuts with Mattery,” “Brown Crunch,” “Cookies and Green.”?
Dr. Lauren Gawne: They’re weird to us because of the semantics of them. They still are English words, or they look like something we’d recognize as English words. I think it’s worth saying artificial intelligence doesn’t know what ice cream is. It’s just using this list of flavors to figure out what kind of patterns could fit into that list.
Dr. Janelle Shane: Exactly. It’s doing it at a very basic level. Like, what kinds of letters tend to come after other letters? What letters are we often finding in combination? Which letters are we never finding in combination? It’ll learn frequent words like “chocolate” or something. It’ll learn how to spell that after some false starts during training, without having any concept of what chocolate is. What it’s trying to do, how it knows it’s making any progress at all, is its job is to try and predict the next letter or the next combination of letters. Then, it just checks its prediction against some example of real texts that it hasn’t seen before that it saved aside to check itself. It’s like a trial and error, guess and check.
03.
Dr. Lauren Gawne: Lauren: So, that’s how it learns “chocolate”? Because it might go in with CH and HC, and every time it goes, “Is HC right? Is HC right?” And the data set is like, “Naw, not really.” But when it’s got the CH for an ice cream list, it’s like, getting lots of positive feedback that that’s gonna appear in “chocolate” and “chip” and “cherry.”
Gretchen McCulloch: It doesn’t have a sense of what’s probable in the world either, right? Because you have some of these flavours like “Peanut Butter Slime,” which those are all English words, it’s just it would make a terrible ice cream because slime and peanut butter and ice cream are not things that go together.
Dr. Janelle Shane: Yeah. Keep in mind, too, the amount of computing power it has to work with is so much less than what it takes for sentience or anything near human level. If you’re looking at raw computing power, the neural nets we have today are somewhere around the level of an earthworm.
04.
Dr. Lauren Gawne: We have things like ice cream names, and you’ve done death metal names, and Halloween costumes, and colours, and that these are all three or four words at most. For ice cream names that have three or four words at most Then, you also have an A.I that was trying to do pick-up lines. Pick-up lines are moving into more of the sentence/couple of sentences-type of thing. As the amount of words you’re trying to generate grows longer, how much more difficult does that make it for the artificial intelligence?
Dr. Janelle Shane: It makes it a lot more difficult. One of the things is that the A.I. I was working with at the time didn’t have very much memory at all. So, it would kind of lose track of things that happened a couple of words ago. It wasn’t really able to figure out then how to make a sentence work or make phrases work. The pick-up lines were definitely a case of, “This is too hard for the A.I.” It struggles, okay, not just the “How do you make a grammatical phrase?” but also “How do you do puns? How do you do innuendo?” These were all things that require a lot of background knowledge that this thing just did not have.
05.
Gretchen McCulloch: Another example that you use in the book is with recipes. It can figure out that you need to list some ingredients, you need to list some instructions, but then those instructions won’t contain the ingredients that were previously mentioned, necessarily, because it doesn’t remember that those are what it listed before.
Dr. Janelle Shane: Yeah. You’ll get something that on the surface at first glance looks like a recipe but not a recipe at all. It’s forgotten its ingredients. It’s copying what humans have written. It’s just going to plough ahead with its best guess at what a human would say.
06.
Gretchen McCulloch: This is where, I think, your famous giraffe question comes from.
Dr. Janelle Shane: It’s a chatbot called Visual Chatbot. It’s designed to answer questions about an image. Its training data is important. In this case, one of the things that they wanted to make sure to avoid was this thing called priming. People tend to ask questions to which the answer tends to be “yes.” They found in an early version of this chatbot that they could get 80% accuracy just by answering “yes” to every single yes-or-no question. They ended up having to hide the image from the person who was asking questions, so that helped a little bit. Now, it’s about 50/50 if you ask a given question whether it’s going to answer yes or no to that. One of the things that they weren’t able to correct was this interesting thing with the giraffes. What happens is, if you ask the question, “How many giraffes do you see?” the chatbot will almost always return a non-zero answer. It is copying how humans tend to answer this question. Humans had not tended to ask the question, “How many giraffes are there?” when they didn’t know if there were any giraffes.
07.
Gretchen McCulloch: Right. You’d say something like, “Are there any giraffes?” The person says, “Yes,” and then you say, “How many giraffes?”
Dr. Janelle Shane: Exactly. If you ask the chatbot, “Are there any giraffes?” it will answer, “No,” quite often. But then, if you follow up with the question, “And how many giraffes do you see?” it’ll say, “Five.”
08.
Dr. Lauren Gawne: One application of this name-generation process you’ve been doing was when you created a list of craft beer names, and a company actually took one of those names to create a beer. Was that a process that you embarked on because you thought this was a good place to experiment with creative naming or how did that come about?
Dr. Janelle Shane: In the case of craft beer names, there’ve actually been companies who have taken each other to court over having beer names that were too close to one another. There’s this need to maybe show there’re ways to still come up with new beer names and we hadn’t exhausted all the possibilities yet. It’s really a collaboration between human and the A.I. where we are curating all of the names that it gives us in order to find the ones that have that perfect balance of following the rules we’ve given it but with a bit of a lateral thinking approach.
09.
Dr. Lauren Gawne: We’ve talked a little bit about how you have to curate the output because it will just keep spitting out silly ice cream names forever. We’ve talked a little bit about some of the problems with the types of data that are put into these processes in terms of, you know, if you don’t set it up very well and you have people answering questions about giraffes in a way that the A.I. is going to implement weirdly. There are bigger and more serious implications for thinking about the kind of data that we are using to create artificial intelligence processes not just with language but particularly for this topic looking at the kinds of data that people use to build artificial intelligence. You talk about this a bit in your book. Where do you see some of the biggest challenges in creating good A.I.?
Dr. Janelle Shane: One of the things is, remember these A.I.s have about the raw computing power of an earthworm, and they don’t have the context, then, to realize that there are some things that the humans do that they probably shouldn’t be copying. Completely unknowingly, they will copy things like racial/gender discrimination and they won’t know that that’s what they’re doing. They won’t know that that’s a bad thing. They just really can’t comprehend it.
10.
Dr. Lauren Gawne: It seems like in these last few years, A.I’s ability to process larger text has gotten better. Is that the case, Janelle?
Dr. Janelle Shane: Yeah, that’s definitely the case. The kinds of things I was doing in 2016 – generating words, short phrases, paint color names, ice cream flavor names, those sorts of things – I wouldn’t think of tackling entire sentences or, let alone, sentences that follow one another that make sense. But now, just pretty much in the last year, there’s been some really big A.I.s that have been trained on millions of pages from the internet. They are much better at generating text. They can generate grammatical sentences most of the time now. Most of the words that they use are real words. They still don’t understand what they’re saying. But I think, yeah, it has gotten better.
往期推荐
Noam Chomsky | Decoding the Human Mind & Neural Nets
本文版权归“理论语言学五道口站”所有，转载请联系本平台。
编辑：安安 黄静雯 董泽扬
排版：黄静雯
审校：董泽扬 吴伟韬 时仲
英文编审负责人：安安 黄静雯