Peter Norvig 的分词问题：如何对里面有拼写错误的词进行分词？

Question

我想了解 Peter Norvig 的拼写校正器是如何工作的。

在他的 jupyter-notebook 标题 here 中，他解释了如何在不使用空格分隔单词的情况下分割字符序列。它工作正常，当顺序中的所有单词都写正确时：

>>> segment("deeplearning")
['deep', 'learning']

但是当序列中的单词（或某些单词）拼写错误时，它会出错：

>>> segment("deeplerning")
['deep', 'l', 'erning']

不幸的是，我不知道如何解决这个问题并使 segment() 函数可以处理拼写错误的连接词。

有人知道如何处理这个问题吗？

Answer 1

It could be achieved by Peter Norvig's algorithm with minor changes. The trick is to add a space character to the alphabet and treat all bigrams separated by space character as a unique word.

Since big.txt doesn't contain deep learning bigram we will have to add a little bit more text to our dictionary. I will use wikipedia library (pip install wikipedia) to get more text.

import re
import wikipedia as wiki
import nltk
from nltk.tokenize import word_tokenize
unigrams =  re.findall(r"\w+", open("big.txt").read().lower())
for deeplerning in wiki.search("Deep Learning"):
    try:
        page = wiki.page(deeplerning).content.lower()
        page = page.encode("ascii", errors="ignore")
        unigrams = unigrams + word_tokenize(page)
    except:
        break

I will create a new dictionary with all unigrams and bigrams:

fo = open("new_dict.txt", "w")
for u in unigrams:
    fo.write(u + "\n")
bigrams = list(nltk.bigrams(unigrams))
for b in bigrams:
    fo.write(" ".join(b)+ "\n")
fo.close()

Now just add a space character to the letters variable in edits1 function, change big.txt to new_dict.txt and change this function:

def words(text): return re.findall(r'\w+', text.lower())

to this:

def words(text): return text.split("\n")

and now correction("deeplerning") returns 'deep learning'!

This trick will perform well if you need a spelling corrector for specific domain. If this domain is big you can try to add to your dictionary only most common unigrams / bigrams.

This question also may help.

Peter Norvig 的分词问题：如何对里面有拼写错误的词进行分词？

Peter Norvig's word segmentation issue: how can I segment words with misspellings inside?

python

nlp

spell-checking

spelling