未來手機能否讀懂你的心思大綱

PATTON OSWALT, an American comedian, once told a story about a text exchange with his girlfriend. “I love you,” she texted. Mr Oswalt began to reply “I love you too.” Only the grouchy comic got as far as “I…” and the predictive texting program began to fill out the text based on his habits. “…hate…”, it provided. Mr Oswalt hit “send” before he could stop himself, and his poor girlfriend ended up receiving a rather offensive text.
帕頓·奧斯瓦爾特是一名美國喜劇演員，他跟觀衆講過自己與女朋友短信交流的故事。“我愛你”，女友在短信中寫道。奧斯瓦爾特先生準備回覆“我也愛你”。只有愛發牢騷的喜劇演員才懂得“我...”的句式效果，這時文字預測功能開始根據他的習慣自動填寫短信，所寫的內容是：“...恨...”。奧斯瓦爾特先生還沒來得及反應，手一抖就按了“發送”鍵，結果他女友收到了這條冒犯意味十足的短信。

Johnson would expect Mr Oswalt's relationships to have moved on from such moments of communicative meltdown. Predictive texting was fairly new in 2009, when Mr Oswalt told his gag. But five years later, smarter and faster processors, plus better wireless broadband, have allowed smartphones to do much better at predicting what users mean, and what they are likely to say next. Apple has implemented new elements of language analysis and prediction in iOS8, its latest mobile operating system, for texting both with thumbs and with speech recognition.
筆者約翰遜希望奧斯瓦爾特先生與女友的關係不要因爲這種溝通失敗而泡湯。2009年，奧斯瓦爾特先生向觀衆講這個噱頭時，文字預測還是比較新奇的功能。但是五年後的今天出現了更快更智能的處理器和更先進的無線寬帶，使智能手機能更好的預測用戶的心思以及下一句想要說什麼。蘋果公司爲最新的手機操作系統IOS8提供了新的語言分析和預測元素，適用於手指打字和語音識別。

Speech recognition relies on a big database of natural human English text. An unclear word can be disambiguated by the words around it, as the software tries to match a string of words to a string in its database. If the computer hears "Four score and seven [mumble] ago", it can scan its database and guess that the missing word is probably "years".
語音識別技術依靠普通人類英語文本大型數據庫。當軟件聽到一連串單詞時，會從數據庫中搜索與之相匹配的語句。所以當軟件遇到一個發音模糊的單詞時，可根據其前後其他單詞來消除歧義。例如，當計算機聽到“四十七[發音模糊]前”，它會搜索數據庫猜測沒聽清的那個單詞可能是“年”。

In iOS7, the user had to dictate an entire passage, hit “done”, and wait for the (usually quite accurate) text to appear after a second or two. In iOS8, though, things get more interesting: the words appear nearly as they are spoken. The real-time appearance of each word allows the curious language pundit to peer into the software’s mind. I tried a deliberate “Four score and seven rrrrs ago.” Each of the first four words appeared almost instantly. After the rrrrs, the system paused for a brief moment, before offering “years ago” on the screen. The software clearly thought “hmm, rrrrs sounds like nothing in my database. But this ‘four score and seven’ is almost always followed by the word ‘years’, and the ‘ago’ seems to prove it.”
在IOS7系統中，用戶必須先口述一整段語音，點擊“完成”鍵，一兩秒後語音被轉化成文字，準確率一般都很高。在IOS8系統中，該功能變得更爲有趣：語音和文字幾乎是同步的。由於每個單詞可以實時出現，因此語言專家可以分析軟件的思維方式。我故意說了“四十七rrrrs前”，前四個單詞中的第一個單詞幾乎立刻出現。當我說到rrrrs時，系統短暫停頓後在屏幕上顯示出“年前”。軟件明顯在想：“嗯...我的數據庫中找不到rrrrs的發音，但“四十七”後面經常出現“年”這個單詞，而“前”這個單詞似乎證明了這一點。

To check my intuition, I tried “rrrrs” for “years” in several situations where "years" is not the obvious word, including a classic line from "Raiders of the Lost Ark": “It’s not the years, honey—it’s the mileage.” The software returned things like “It’s not the errors honey it’s the mileage,” and “It’s not Thursday honey it’s the mileage.” The line obviously isn’t in Apple’s training text as frequently as “Four score and seven years ago.”
爲了驗證我的直覺判斷，我在多個情景中將“年（years）”說成了“rrrrs”，在這些情景中單詞“年（years）”並不是顯而易見就能推測出來的，包括《奪寶奇兵》中的經典臺詞：“關鍵不在於車齡，寶貝，而在於行駛里程”。這時軟件顯示出的文字是：“關鍵不在於錯誤（errors），寶貝，而在於行駛里程”，以及 “關鍵不在於週四（Thursday），寶貝，而在於行駛里程”。這句話在蘋果數據庫中出現的頻率明顯不如 “四十七年前”。

Another much-touted advance in iOS8 is predictive text for typing. When composing a text or e-mail, users see three words above the keyboard at any given time—the three words Apple reckons they are most likely to use next. As with speech recognition, the software must first be trained on a bunch of actual English text. But Apple also claims that the software learns from each individual user over time.
蘋果IOS8系統中另一個備受追捧的進步是打字預測功能。當用戶在編輯文字或電子郵件時，隨時能看到鍵盤上會浮現出三個單詞，蘋果認爲這三個單詞是你接下來最有可能用到的。與語音識別同理，這款軟件也必須事先接受真實英語文本的訓練。但蘋果也表示該軟件隨着時間也會學習單個用戶的使用習慣。iOS8 has had about a month to learn your columnist’s habits. At present, the three words Apple thinks Johnson is mostly to use to start a text are “I”, “The” and “I’m”. If that sounds depressingly ego-centric, remember that “I” is the most common word in spoken English. Pressing “I” reveals the three words most likely to follow “I”: “love”, “don’t” and “just”. Johnson is apparently not as disagreeable as Mr Oswalt.
IOS8系統用了一個月時間學習筆者的習慣。現在蘋果認爲筆者在編輯文字時，開頭最常用的三個單詞是“我”，“這”，“我是”。儘管這種以自我爲中心的感覺令人鬱悶，但請記住“我”是英文口語中最常用的單詞。在鍵盤上點擊“我”會顯示接下來最有可能用到的三個單詞：“愛”，“不”，“僅僅”。筆者顯然不像奧斯瓦爾特先生那般不友好。

What is striking, however, is that though the software can pick words that are likely to follow the previous word, the trick does not produce great phrases. Repeatedly pressing the middle of the three choices on my phone results in
令人驚奇的是，該軟件能從先前出現過的預測結果中選詞，但這種技巧無法產生令人滿意的短語。在我鍵盤上重複重複點擊第二個備選單詞，結果產生下面的文本：

“The day I have a great way of the year and the other hand is the only thing that would have to go back and I don’t think that I have a great way of life and the day I have to go back and the other hand…”
我以一年中良好的方式度過了這一天（The day I have a great way of the year），另一方面（the other hand），這一天是我必須要回顧的（have to go back），我認爲自己沒有良好的生活方式（I have a great way of life），這一天是我必須要回顧的（have to go back），另一方面（the other hand）...

Curiously, sometimes the system repeats its own predictions (“and the other hand” occurs twice). But sometimes it doesn’t (“I have a great way of” is once followed by “the year”, and the second time followed by “of life”). And if I clear the whole mess and begin again, repeatedly tapping the same button gives me a different string. What is going on behind the scenes is unclear. (Apple did not respond to requests to clarify.)
令人驚奇的是，系統有時會重複之前的預測結果（“and the other hand”出現過兩次）。但有時又不會重複（第一次出現的“I have a great way of”後面跟隨的是“the year”，而第二次出現時後面跟隨的是“of life”）。如果我刪掉這些亂七八糟的文字重新編輯，然後同樣重複點擊第二個備選單詞，產生的一連串單詞會與上次不同。這究竟是怎麼一回事我不得而知，蘋果公司對於澄清此事的請求未作迴應。

When this new predictive feature was announced, a few observers harrumphed: offering the next potential word allows a writer to skip the work of choosing words. Well, to choose just one word: nonsense. At the most, iOS8 will allow you to avoid typing some fairly long but frequent words. But it’s not ready to appease your boss, apologise to your spouse or do your homework for you. In 1965 Michael Frayn imagined automating journalism in his novel The Tin Men, automatically generating catchy headlines and composing predictable stories (like ones on the royal family) with a string of clichés. Fortunately, half a century later, we hacks must still be paid to string out clichés by hand.
當這種新潮的文字預測功能剛問世時，少數觀察人士曾對此嗤之以鼻：預測下一個可能用到的單詞會讓寫字的人越過選詞過程。只能用一個詞來形容：荒唐。IOS8系統最多就是讓你不必重複寫出冗長又經常使用的單詞，但它沒想代替你去安撫你的老闆，向你的愛人道歉，或者幫你做作業。1965年，邁克爾·弗萊恩曾在他的小說《鐵皮人》中暢想過新聞工作自動化的情景：自動產生朗朗上口的新聞標題；編寫可預測的故事（比如皇室故事）；一連串的陳詞濫調。幸運的是，半個世紀後的今天，我們這些拿工資的僱傭文人仍需親手編寫陳詞濫調。