文章來源：《經(jīng)濟學(xué)人》Apr 29th 2023 期 Culture 欄目 The language instinct

[Paragraph 1]

When deep blue, a chess computer, defeated Garry Kasparov, a world champion, in 1997 many gasped in fear of machines triumphing over mankind.?In the intervening years, artificial intelligence has done some astonishing things, but none has managed to capture the public imagination in quite the same way.?Now, though, the astonishment of the Deep Blue moment is back, because computers are employing something that humans consider their defining ability: language.

1997年，當(dāng)國際象棋計算機 "深藍(lán) "擊敗世界冠軍加里-卡斯帕羅夫時，許多人都因機器戰(zhàn)勝人類而感到恐懼。在此后的幾年里，人工智能做了一些驚人的事情，但沒有一件事能以完全相同的方式吸引公眾的注意力.不過現(xiàn)在，深藍(lán)時刻的震撼場面再次出現(xiàn)了，因為計算機正在運用人類認(rèn)為是其決定性能力的東西：語言。

[Paragraph 2]

Or are they? Certainly, large language models (LLMs), of which the most famous is ChatGPT, produce what looks like?impeccable?human writing.?But a debate has ensued about what the machines are actually doing internally, what it is that humans, in turn, do when they speak—and, inside the academy, about the theories of the world’s most famous linguist, Noam Chomsky.

或者說他們是這樣嗎？當(dāng)然，大型語言模型中最著名的是ChatGPT，它可以產(chǎn)生似乎無可挑剔的人類寫作。但是，關(guān)于機器內(nèi)部實際做什么、人類在說話時又做什么的爭論隨之而來——在學(xué)術(shù)界，關(guān)于世界上最著名的語言學(xué)家諾姆·喬姆斯基理論的爭論非?；馃帷?/p>

[Paragraph 3]

Although Professor Chomsky’s ideas have changed considerably since he rose to?prominence?in the 1950s, several elements have remained fairly constant.?He and his followers argue that human language is different in kind (not just degree of expressiveness) from all other kinds of communication.?All human languages are more similar to each other than they are to, say, whale song or computer code.?Professor Chomsky has frequently said a Martian visitor would conclude that all humans speak the same language, with surface variation.

1950年代，喬姆斯基教授的思想聲名顯赫，雖然現(xiàn)在已經(jīng)發(fā)生了很大變化，但有幾個觀點一直保持不變。他和他的追隨者認(rèn)為，人類語言在種類上（不僅僅是表達(dá)能力的程度）與所有其他種類的交流不同。所有人類語言彼此之間具有相似性，且比鯨魚的歌聲或計算機代碼之間的相似性要高。喬姆斯基教授經(jīng)常說，一個來自火星的訪客可能會得出結(jié)論，所有人類都說同一種語言，只是表面有所不同。

[Paragraph 4]

Perhaps most notably, Chomskyan theories hold that children learn their native languages with astonishing speed and ease despite “the poverty of the stimulus”: the sloppy and occasional language they hear in childhood.?The only explanation for this can be that some kind of?predisposition?for language is built into the human brain.

或許最值得注意的是，喬姆斯基的理論認(rèn)為，盡管 "刺激的貧乏"（即兒童在童年時期聽到的馬虎和偶然的語言），但孩子們還是以驚人的速度又輕松地學(xué)會了他們的母語。對此的唯一解釋是，人類大腦中內(nèi)置了某種語言功能。

[Paragraph 5]

Chomskyan ideas have dominated the linguistic field of syntax since their birth.?But many linguists are?strident?anti-Chomskyans. And some are now seizing on the capacities of LLMs to attack Chomskyan theories anew.

喬姆斯基的思想自誕生以來一直主導(dǎo)著句法語言學(xué)領(lǐng)域。但許多語言學(xué)家都是堅定的反喬姆斯基主義者。而且有些人現(xiàn)在正利用大型語言模型的能力，重新攻擊喬姆斯基的理論。

[Paragraph 6]

Grammar has a?hierarchical, nested structure involving units within other units. Words form phrases, which form clauses, which form sentences and so on.?Chomskyan theory posits a mental operation, “Merge”, which glues smaller units together to form larger ones that can then be operated on further (and so on).?In a recent New York Times?op-ed, the man himself (now 94) and two co-authors said “we know” that computers do not think or use language as humans do, referring?implicitly?to this kind of cognition.?LLMs, in effect, merely predict the next word in a string of words.

語法具有層次結(jié)構(gòu)、嵌套結(jié)構(gòu)，涉及其他單元中的單元。單詞構(gòu)成短語，短語構(gòu)成從句，句子構(gòu)成句子等等。喬姆斯基理論提出了一種心理操作，即 "合并"，它將較小的單元粘在一起形成較大的單元，然后可以進(jìn)一步操作（等等）。在最近的《紐約時報》專欄文章中，他本人（現(xiàn)年94歲）和兩位合著者說，"我們知道 "計算機不像人類那樣思考或使用語言，暗指這種認(rèn)知攻擊。實際上，大型語言模型只是預(yù)測一連串單詞中的下一個單詞。

[Paragraph 7]

Yet it is hard, for several reasons, to?fathom?what LLMs “think”.?Details of the programming and training data of commercial ones like ChatGPT are?proprietary.?And not even the programmers know exactly what is going on inside.

然而，由于一些原因，我們很難理解大型語言模型的 "想法"。ChatGPT等商業(yè)軟件的編程和訓(xùn)練數(shù)據(jù)的細(xì)節(jié)是有私有的。甚至連程序員都不知道里面到底發(fā)生了什么。

[Paragraph 8]

Linguists have, however, found clever ways to test LLMs’ underlying knowledge, in effect tricking them with probing tests.?And indeed, LLMs seem to learn nested, hierarchical grammatical structures, even though they are exposed to only linear input, ie, strings of text.?They can handle novel words and grasp parts of speech.?Tell ChatGPT that “dax” is a verb meaning to eat a slice of pizza by folding it, and the system deploys it easily: “After a long day at work, I like to relax and dax on a slice of pizza while watching my favourite TV show.” (The imitative element can be seen in “dax on”, which ChatGPT probably patterned on the likes of “chew on” or “munch on”.)

然而，語言學(xué)家找到了測試大型語言模型基礎(chǔ)知識的巧妙方法，實際上是用探測性測試來誤導(dǎo)它們。事實上，大型語言模型似乎學(xué)習(xí)了嵌套的、層次化的語法結(jié)構(gòu)，盡管它們只接觸到線性輸入，即文本字符串。它們可以處理新詞并掌握詞性。例如告訴ChatGPT，"dax "是一個動詞，意思是折疊著吃一片比薩餅，系統(tǒng)就能輕松地處理它："在漫長的一天工作之后，我喜歡去放松放松，一邊看著我最喜歡的電視節(jié)目，一邊吃著比薩餅。" (在 "dax on "中可以看到模仿的成分，ChatGPT可能是按照 "chew on "或 "munch on "詞組結(jié)構(gòu)。)

[Paragraph 9]

What about the “poverty of the stimulus”??After all, GPT-3 (the LLM underlying ChatGPT until the recent release of GPT-4) is estimated to be trained on about 1,000 times the data a human ten-year-old is exposed to.?That leaves open the possibility that children have an inborn tendency to grammar, making them far more proficient than any LLM.?In a?forthcoming?paper in Linguistic Inquiry, researchers claim to have trained an LLM on no more text than a human child is exposed to, finding that it can use even rare bits of grammar.?But other researchers have tried to train an LLM on a database of only child-directed language (that is, of transcripts of carers speaking to children).?Here LLMs fare far worse. Perhaps the brain really is built for language, as Professor Chomsky says.

那么 "刺激的貧乏 "怎么解釋？畢竟，GPT-3（這是ChatGPT的基礎(chǔ)大型語言模型，最近發(fā)布了GPT-4）所接受的訓(xùn)練數(shù)據(jù)大約是 10 歲兒童所接觸數(shù)據(jù)的 1,000 倍。這就留下了一種可能性，即兒童天生就有語法功能使他們比任何大型語言模型都熟練得多。在《語言學(xué)探索》上即將發(fā)表的一篇論文中，研究人員聲稱他們訓(xùn)練的大型語言模型所用的文本不超過人類兒童接觸到的文本，發(fā)現(xiàn)它甚至可以使用罕見的語法。但是其他研究人員已嘗試用僅包含面向兒童的語言（也就是看護者與兒童對話的記錄）的數(shù)據(jù)庫訓(xùn)練大型語言模型。這種情況下，大型語言模型的表現(xiàn)要差得多。正如喬姆斯基教授所說，也許大腦真的是為語言而生的。

[Paragraph 10]

It is difficult to judge. Both sides of the argument are?marshalling?LLMs to make their case.?The?eponymous?founder of his school of linguistics has offered only a?brusque riposte.?For his theories to survive this challenge, his camp will have to put up a stronger defence.

這很難判斷。爭論雙方都在用大型語言模型來證明自己的觀點。這一語言學(xué)派的同名創(chuàng)始人（喬姆斯基）只是簡單直接地反駁了一下。為了使他的理論在這次挑戰(zhàn)中站住腳，他的陣營必須提出更有力的觀點來辯護。