如何比较不完全匹配的字符串

Question

I need to compare two output strings namely the original transcription and a transcription of a Speech-to-Text service. Often numbers are written in a numerical format or as a word eg "four" or "4". How to compare strings considering these different methods of transcribing?

So far I just transformed both strings in lower case letters and split each word with a space as seperator.

#Read the two files and store them in s1_raw and s2_raw
with open('original.txt', 'r') as f:
    s1_raw = f.read()
with open('comparison.txt', 'r') as f:
    s2_raw = f.read()

#Transform all letters to minuscule letter
s1 = s1_raw.lower()
s2 = s2_raw.lower()

#Split texts with space as seperator to have a list of words
s1_set = s1.split(' ')
s2_set = s2.split(' ')

#Used later for confidence calculation
count1 = len(s1_set)
count2 = 0
x = 0

#Check which string is longer to prevent running out of indices
if len(s1_set) < len(s2_set):
    #Loop through whole list and compare word by word
    for x in range (0, len(s1_set)):
        if s1_set[x] == s2_set[x]:
            count2 += 1
        x += 1
else:
    #Loop through whole list and compare word by word
    for x in range (0, len(s2_set)):
        if s1_set[x] == s2_set[x]:
            count2 += 1
        x += 1

#Confidence level= correct words divided by total words
confidence = count2/count1

#Print out result
print('The confidence level of this service is {:.2f}%'.format(confidence*100))

I want to measure the accuracy of the transcription for several *.txt files and consider all the different ways of how the different Speech-to-Text services transcribe.

Answer 1

You have to normalize the text before comparing it. First decide if four or 4 is your canonical form and convert all strings to that form.

For example, if four is the canonical form, then write code to replace 1 with one , 213 with two hundred and thirteen , and so on, and do the comparison with these.

Actually, I think that it is better to normalize to 4 rather than four since there can be more than one way to express a number in some languages. By preferring 4 it is possible to normalize all equivalent transcriptions to one single form.

Answer 2

Thanks @Michael Veksler. I tried the NLTK library now to split the string into word lists more efficiently. Also, I tried to look for synonyms of each word and compare if the synonyms match. This still doesn't really solve the task so I wonder what else I could try.

I use those two libraries:

from nltk.tokenize import word_tokenize
from nltk.corpus import wordnet

Splitting the words is as easy as:

s1_set = word_tokenize(list1)

Now I try to find synonyms of the words and take the first found synonym. I append it to an empty list named "wl1". I check before, if any synonym is found, since this is not always the case.

for i in range(0, (len(s1_set)-1)):
    #Find synonym of word in s1_set index i
    t1 = wordnet.synsets(s1_set[i])
    #Ensure t1 isn't empty
    if t1:
        wl1.append(t1[0].lemmas()[0].name())

Then I am again comparing word by word like in my first post above. This method isn't a satisfying solution to my problem either. Can anyone think of a better method?