Artificial Intelligence Has a Problem with Grammar人工智能遭遇語法問題

萊恩·格林 譯/傅穎 Lane Greene

The hitch illuminates the nature of language.

這一難題揭示了語言的本質。

If you frequently Google language-related questions， whether out of interest or need， youve probably seen an advertisement for Grammarly， an automated grammar-checker. In ubiquitous YouTube spots Grammarly touts its ability not only to fix mistakes， but to improve style and polish too. Over more than a decade it has sprawled into many applications： it can check emails， phone messages or longer texts composed in Microsoft Word and Google Docs， among other formats.

Does it achieve what it purports to？ Sometimes. But sometimes Grammarly doesnt do what it should， and sometimes it even does what it shouldnt. These strengths and failings hint at the essence of language and the peculiarity of human intelligence， as opposed to the artificial sort as it stands today.

Begin with the strengths. In a rough piece of student writing， Johnson counted 14 errors. Grammarly flagged five. For example， it sensibly suggested inserting a hyphen in “post cold war [world]”. It spotted a missing “the” in the phrase “with [the] European economy”. And it noticed an absent “about” in “wondering [about] the state of Europe”. By using Grammarly， the author of this essay could have avoided some red ink.

On the other hand， Grammarly has a problem with false positives， calling out mistakes that are not. The other two suggestions were not disastrous， but neither did they relate to “critical errors” as Grammarly maintains. In the assertion that enlargement had “created a fatigue” within the European Union， Grammarly needlessly suggested deleting the “a”. In another error-ridden sentence it recommended removing a comma， which fixed none of the problems. This false-positive tendency is not a deal-breaker for reasonably skilled writers who just want a second pair of eyes; you can dismiss any suggestion you like. But truly struggling scribblers might not know when Grammarlys ideas would make their prose worse rather than better.

Then there are the false negatives， or the mistakes Grammarly fails to notice. Depending on the text， Grammarly can seem to miss more errors than it marks. The companys chief executive， Brad Hoover， describes it as a “coach， not a crutch”—which sets expectations more appropriately than some of the ads do.

Artificial-intelligence systems like Grammarly are trained with data; for instance， translation software is fed sentences translated by humans. Grammarlys training data involve a large number of standard error-free sentences （so it knows what good English should look like） and human-corrected sentences （so the software can find the patterns of fixes that human editors might make）. Developers also manually add certain rules to the patterns Grammarly has taught itself. The software then looks at a users prose： if a string of words seems ungrammatical， it tries to spot how the putative mistake most closely resembles one from its training inputs.

All this shows how far artificial “intelligence” is from the human kind （which Grammarly wants to correct to “humankind”）. Computers outpace humans at problems that can be cracked with pure maths， such as chess. Advances in language technology have been impressive in， for example， speech recognition， which involves another sort of statistical guess—whether or not a stretch of sound matches a certain string of words. One Grammarly feature that works fairly well is sentiment analysis. It can rate the tone of an email before you send it， after being trained on texts that have been assessed by humans， for example as “admiring” or “confident”.

But grammar is the real magic of language， binding words into structures， binding those structures into sentences， and doing so in a way that maps onto meaning. And at this crucial structure-meaning interface， machines are no match for humans. Computers can parse （grammatical） sentences fairly well， labelling things like nouns and verb phrases. But they struggle with sentences that are difficult to analyse， precisely because they are ungrammatical—in other words， written by the kind of person who needs Grammarly.

To correct such prose requires knowing what the writer intended. But computers dont work in meaning or intention; they work in formulae. Humans， by contrast， can usually understand even rather mangled syntax， because of the ability to guess the contents of other minds. Grammar-checking computers illustrate not how bad humans are with language， but just how good.

如果你經常上谷歌搜索與語言相關的問題，無論是出于興趣還是出于需要，你都可能看到過Grammarly的廣告，這是一款自動語法檢查工具。在漫天的優兔插播廣告中，Grammarly宣稱它不僅能夠糾正錯誤，還能改進文風，給文章潤色。10多年來，它已經打入許多應用程序：它能夠檢查電子郵件、手機短信，或是以微軟Word文檔、谷歌文檔等其他格式編寫的長文本。

那它說到做到了嗎？有時候做到了。但有時候Grammarly失職了，有時候它甚至做了不該做的。這些優缺點暗示出語言的本質以及人類智能的特性，而非當今所謂人工智能的特點。

先說Grammarly的優點。在一篇質量不高的學生作文中，《經濟學人》的約翰遜語言專欄標出了14處錯誤。Grammarly則標記了5處。例如，它建議在詞組“post cold war [world]（后冷戰[世界]）”中插入連字符，這很合理;它發現，短語“with [the] European economy（歐洲經濟）”漏了the;它還注意到，“wondering [about] the state of Europe（對歐洲狀況的思考）”少了about。借助Grammarly，這篇文章的作者可以避免一些錯誤。

而另一方面，Grammarly存在誤報問題，它會指出并非錯誤的錯誤。Grammarly給出的另外兩條建議雖不至于離譜，但也談不上它所認為的“嚴重錯誤”。針對歐盟擴大在內部“created a fatigue（引發了疲勞）”這句話，Grammarly建議刪除a，這多此一舉。另一個滿是錯誤的句子則被建議刪除逗號，可這并未解決任何問題。對那些只想多一雙眼睛檢查的寫作高手來說，這種頻現的誤報并不會壞事：你可以忽略想忽略的任何建議。但那些絞盡腦汁、水平不高的作者可能無法判斷，在什么情況下Grammarly的建議會幫倒忙。

此外，Grammarly還存在漏報問題，即無法發現某些錯誤。Grammarly漏掉的錯誤可能比標記出來的還要多，視文本內容而定。該公司首席執行官布拉德·胡佛將Grammarly形容為“教練，而非拐杖”。相較一些廣告，這個比方更為恰當地設定了此款軟件該符合的期望。

像Grammarly這樣的人工智能系統是用數據訓練的。例如，翻譯軟件的訓練數據是人工翻譯的句子。Grammarly的訓練數據包括大量標準無誤的句子（所以它知道好的英語應該是什么樣子）和人工糾正的句子（所以它能發覺人工編輯可能采取的改錯模式）。開發人員還將某些規則手動添加到Grammarly的自學修改模式中。這樣，當該軟件檢查用戶文章時，如果一串單詞看起來不合語法，它便會試圖找出假定的錯誤與訓練輸入的錯誤最相似的地方。

所有這些表明，人工“智能”和人的智能[即human kind，Grammarly會把這個詞組改為“humankind（人類）”]相去甚遠。計算機在下國際象棋等純數學問題上比人厲害。它在語言技術方面的進步也令人贊嘆，比如語音識別，這涉及另一種統計猜測，即一段聲音與某串單詞是否匹配。Grammarly具備一項很棒的功能：情緒分析。它可以在電子郵件發送之前對其語氣進行評估。它接受過訓練，見識過哪些文本被人類評定為“贊賞的”或“自信的”等等。

然而，語言真正的神奇之處在于語法，它將單詞綁定到結構中，將這些結構綁定到句子中，使之表情達意。結構與意義之間的交互至關重要，在這點上，機器無法與人類相比。盡管計算機能很好地（從語法上）解析句子，標出諸如名詞和動詞短語等句子成分，但面對難以分析的句子，計算機束手無策，這恰恰是因為這些句子不符合語法，換句話說，寫出這些句子的正是需要Grammarly的人。

要修改這類文本，就要知道作者的意圖。但是，計算機無法理解意義或意圖，它們靠的是公式。相比之下，人類因為有能力猜測別人的想法，所以通常能夠理解十分混亂的句法。用計算機檢查語法，并不能說明人類處理語言的能力有多么糟糕，相反，這只能說明人類的語言能力十分出色。

（譯者為“《英語世界》杯”翻譯大賽獲獎者）