Facebook 修正錯字的新演算法

先前 Facebook 已經先發表過 fastText 了,在這個月的月初又發表了另外一個演算法 Misspelling Oblivious Embeddings (MOE),是搭著本來的 fastText 而得到的改善:「A new model for word embeddings that are resilient to misspellings」。

Facebook 的說明提到在 user-generated text 的內容上,MOE 的效果比 fastText 好:

We checked the effectiveness of this approach considering different intrinsic and extrinsic tasks, and found that MOE outperforms fastText for user-generated text.

論文發表在 arXiv 上:「Misspelling Oblivious Word Embeddings」。

依照介紹,fastText 的重點在於 semantic loss,而 MOE 則多了 spell correction loss:

The loss function of fastText aims to more closely embed words that occur in the same context. We call this semantic loss. In addition to the semantic loss, MOE also considers an additional supervisedloss that we call spell correction loss. The spell correction loss aims to embed misspellings close to their correct versions by minimizing the weighted sum of semantic loss and spell correction loss.

不過目前 GitHub 上的 facebookresearch/moe 只有放 dataset,沒有 open source 出來讓人直接用,可能得自己刻...


2013 年時在「Automatic Translation Without Dictionaries」這邊看到關於機器翻譯時的自我學習方式,裡面提到了「How Google Converted Language Translation Into a Problem of Vector Space Mathematics」這篇報導,而裡面提到的論文則是 Google 發表在 arXiv 上的「Exploiting Similarities among Languages for Machine Translation」這篇。

最近看到「The Illustrated Word2vec」這篇,把五年多前的記錄交叉拉出來看... 這個算式算是給了大家基本的想法,透過公式來解釋文字的意義:

拉出這樣的關係後,就有機會學習新的詞彙... 進而用在其他語言的翻譯上。


是一篇老文章了... (2014 年的文章,最近從其他地方提起)

這邊講的是英文,不過同樣方式也可以拿來分析其他語言:「The distribution of letters in English words」,原始文章在「Graphing the distribution of English letters towards the beginning, middle or end of words」。


The data is from the entire Brown corpus in the Natural Language Toolkit. It's a smaller and out-of-date corpus, but it's open source and easy to obtain. I repeated the analysis with COHA, the Corpus of Historical American English, a well-curated, proprietary data set from Brigham Young University for which I have a license, and the only differences were in rare letters like "z" or "x".


在「The future of education is plain text」這邊看到對純文字格式的推廣 XDDD


  • Plain text is always compatible
  • Plain text is easy to mix and match
  • Plain text is easy to maintain
  • Plain text is lightweight
  • Plain text is always forward compatible


這讓我想起 George R. R. MartinConan 上提到他是用 WordStar 4.0 寫《冰與火之歌》的採訪片段:

印 "#" 比印 "B" 來的快的問題

這篇是兩年前在 StackOverflow 上的問題:「Why is printing “B” dramatically slower than printing “#”?」。

問問題的人這段程式跑了 8.52 秒:

Random r = new Random();
for (int i = 0; i < 1000; i++) {
    for (int j = 0; j < 1000; j++) {
        if(r.nextInt(4) == 0) {
        } else {


而把上面的 # 換成 B 就變成 259.152 秒。

答案是與 word-wrapping 有關:

Pure speculation is that you're using a terminal that attempts to do word-wrapping rather than character-wrapping, and treats B as a word character but # as a non-word character. So when it reaches the end of a line and searches for a place to break the line, it sees a # almost immediately and happily breaks there; whereas with the B, it has to keep searching for longer, and may have more text to wrap (which may be expensive on some terminals, e.g., outputting backspaces, then outputting spaces to overwrite the letters being wrapped).

But that's pure speculation.

這真是細節 XDDD