Fast n-gram calculation

问题:

I'm using NLTK to search for n-grams in a corpus but it's taking a very long time in some cases. I've noticed calculating n-grams isn't an uncommon feature in other packages (apparently Haystack has some functionality for it). Does this mean there's a potentially faster way of finding n-grams in my corpus if I abandon NLTK? If so, what can I use to speed things up?

回答1:

Since you didn't indicate whether you want word or character-level n-grams, I'm just going to assume the former, without loss of generality.

I also assume you start with a list of tokens, represented by strings. What you can easily do is write n-gram extraction yourself.

def ngrams(tokens, MIN_N, MAX_N):
    n_tokens = len(tokens)
    for i in xrange(n_tokens):
        for j in xrange(i+MIN_N, min(n_tokens, i+MAX_N)+1):
            yield tokens[i:j]

Then replace the yield with the actual action you want to take on each n-gram (add it to a dict, store it in a database, whatever) to get rid of the generator overhead.

Finally, if it's really not fast enough, convert the above to Cython and compile it. Example using a defaultdict instead of yield:

def ngrams(tokens, int MIN_N, int MAX_N):
    cdef Py_ssize_t i, j, n_tokens

    count = defaultdict(int)

    join_spaces = " ".join

    n_tokens = len(tokens)
    for i in xrange(n_tokens):
        for j in xrange(i+MIN_N, min(n_tokens, i+MAX_N)+1):
            count[join_spaces(tokens[i:j])] += 1

    return count

回答2:

You might find a pythonic, elegant and fast ngram generation function using zip and splat (*) operator here :

def find_ngrams(input_list, n):
  return zip(*[input_list[i:] for i in range(n)])

回答3:

For character-level n-grams you could use the following function

def ngrams(text, n):
    n-=1
    return [text[i-n:i+1] for i,char in enumerate(text)][n:]