Discovering Poetic Form with NLTK and CMU Dict

Question

Edit: This code has been worked on and released as a basic module: https://github.com/hyperreality/Poetry-Tools

I'm a linguist who has recently picked up python and I'm working on a project which hopes to automatically analyze poems, including detecting the form of the poem. I.e. if it found a 10 syllable line with 0101010101 stress pattern, it would declare that it's iambic pentameter. A poem with 5-7-5 syllable pattern would be a haiku.

I'm using the following code, part of a larger script, but I have a number of problems which are listed below the program:

corpus in the script is simply the raw text input of the poem.

import sys, getopt, nltk, re, string
from nltk.tokenize import RegexpTokenizer
from nltk.util import bigrams, trigrams
from nltk.corpus import cmudict
from curses.ascii import isdigit

...

def cmuform():
    tokens = [word for sent in nltk.sent_tokenize(corpus) for word in nltk.word_tokenize(sent)]
    d = cmudict.dict()
    text = nltk.Text(tokens)
    words = [w.lower() for w in text]
    regexp = "[A-Za-z]+"
    exp = re.compile(regexp)

    def nsyl(word):
        lowercase = word.lower()
        if lowercase not in d:
                return 0
        else:
            first = [' '.join([str(c) for c in lst]) for lst in max(d[lowercase])]
            second = ''.join(first)
            third = ''.join([i for i in second if i.isdigit()]).replace('2', '1')
            return third 
                #return max([len([y for y in x if isdigit(y[-1])]) for x in d[lowercase]])      

    sum1 = 0
    for a in words:
            if exp.match(a):
            print a,nsyl(a),
                sum1 = sum1 + len(str(nsyl(a)))

    print "\nTotal syllables:",sum1

I guess that the output that I want would be like this:

1101111101

0101111001

1101010111

The first problem is that I lost the line breaks during the tokenization, and I really need the line breaks to be able to identify form. This should not be too hard to deal with though. The bigger problems are that:

• I can't deal with non-dictionary words. At the moment I return 0 for them, but this will confound any attempt to identify the poem, as the syllabic count of the line will probably decrease.
• In addition, the CMU dictionary often says that there is stress on a word - '1' - when there is not - '0 - . Which is why the output looks like this: 1101111101, when it should be the stress of iambic pentameter: 0101010101
  So how would I add some fudging factor so the poem still gets identified as iambic pentameter when it only approximates the pattern? It's no good to code a function that identifies lines of 01's when the CMU dictionary is not going to output such a clean result. I suppose I'm asking how to code a 'partial match' algorithm.

Answer 1

Welcome to stack overflow. I'm not that familiar with Python, but I see you have not received many answers yet so I'll try to help you with your queries.

First some advice: You'll find that if you focus your questions your chances of getting answers are greatly improved. Your post is too long and contains several different questions, so it is beyond the "attention span" of most people answering questions here.

Back on topic:

Before you revised your question you asked how to make it less messy. That's a big question, but you might want to use the top-down procedural approach and break your code into functional units:

1. split corpus into lines
2. For each line: find the syllable length and stress pattern.
3. Classify stress patterns.

You'll find that the first step is a single function call in python:

corpus.split("\n");

and can remain in the main function but the second step would be better placed in its own function and the third step would require to be split up itself, and would probably be better tackled with an object oriented approach. If you're in academy you might be able to convince the CS faculty to lend you a post-grad for a couple of months and help you instead of some workshop requirement.

Now to your other questions:

Not loosing line breaks: as @ykaganovich mentioned, you probably want to split the corpus into lines and feed those to the tokenizer.

Words not in dictionary/errors: The CMU dictionary home page says:

Find an error? Please contact the developers. We will look at the problem and improve the dictionary. (See at bottom for contact information.)

There is probably a way to add custom words to the dictionary / change existing ones, look in their site, or contact the dictionary maintainers directly. You can also ask here in a separate question if you can't figure it out. There's bound to be someone in stackoverflow that knows the answer or can point you to the correct resource. Whatever you decide, you'll want to contact the maintainers and offer them any extra words and corrections anyway to improve the dictionary.

Classifying input corpus when it doesn't exactly match the pattern: You might want to look at the link ykaganovich provided for fuzzy string comparisons. Some algorithms to look for:

• Levenshtein distance: gives you a measure of how different two strings are as the number of changes needed to turn one string into another. Pros: easy to implement, Cons: not normalized, a score of 2 means a good match for a pattern of length 20 but a bad match for a pattern of length 3.
• Jaro-Winkler string similarity measure: similar to Levenshtein, but based on how many character sequences appear in the same order in both strings. It is a bit harder to implement but gives you normalized values (0.0 - completely different, 1.0 - the same) and is suitable for classifying the stress patterns. A CS postgrad or last year undergrad should not have too much trouble with it ( hint hint ).

I think those were all your questions. Hope this helps a bit.

Answer 2

To preserve newlines, parse line by line before sending each line to the cmu parser.

For dealing with single-syllable words, you probably want to try both 0 and 1 for it when nltk returns 1 (looks like nltk already returns 0 for some words that would never get stressed, like "the"). So, you'll end up with multiple permutations: 1101111101 0101010101 1101010101

and so forth. Then you have to pick ones that look like a known forms.

For non-dictionary words, I'd also fudge it the same way: figure out the number of syllables (the dumbest way would be by counting the vowels), and permutate all possible stresses. Maybe add some more rules like "ea is a single syllable, trailing e is silent"...

I've never worked with other kinds of fuzzying, but you can check https://stackoverflow.com/questions/682367/good-python-modules-for-fuzzy-string-comparison for some ideas.

Answer 3

This is my first post on stackoverflow. And I'm a python newbie, so please excuse any deficits in code style. But I too am attempting to extract accurate metre from poems. And the code included in this question helped me, so I post what I came up with that builds on that foundation. It is one way to extract the stress as a single string, correct with a 'fudging factor' for the cmudict bias, and not lose words that are not in the cmudict.

import nltk
from nltk.corpus import cmudict

prondict = cmudict.dict()

#
# parseStressOfLine(line) 
# function that takes a line
# parses it for stress
# corrects the cmudict bias toward 1
# and returns two strings 
#
# 'stress' in form '0101*,*110110'
#   -- 'stress' also returns words not in cmudict '0101*,*1*zeon*10110'
# 'stress_no_punct' in form '0101110110'


def parseStressOfLine(line):

    stress=""
    stress_no_punct=""
    print line

    tokens = [words.lower() for words in nltk.word_tokenize(line)] 
    for word in tokens:        

        word_punct =  strip_punctuation_stressed(word.lower())
        word = word_punct['word']
        punct = word_punct['punct']

        #print word

        if word not in prondict:
            # if word is not in dictionary
            # add it to the string that includes punctuation
            stress= stress+"*"+word+"*"
        else:
            zero_bool=True
            for s in prondict[word]:
                # oppose the cmudict bias toward 1
                # search for a zero in array returned from prondict
                # if it exists use it
                # print strip_letters(s),word
                if strip_letters(s)=="0":
                    stress = stress + "0"
                    stress_no_punct = stress_no_punct + "0"
                    zero_bool=False
                    break

            if zero_bool:
                stress = stress + strip_letters(prondict[word][0])
                stress_no_punct=stress_no_punct + strip_letters(prondict[word][0])

        if len(punct)>0:
            stress= stress+"*"+punct+"*"

    return {'stress':stress,'stress_no_punct':stress_no_punct}



# STRIP PUNCTUATION but keep it
def strip_punctuation_stressed(word):
    # define punctuations
    punctuations = '!()-[]{};:"\,<>./?@#$%^&*_~'
    my_str = word

    # remove punctuations from the string
    no_punct = ""
    punct=""
    for char in my_str:
        if char not in punctuations:
            no_punct = no_punct + char
        else:
            punct = punct+char

    return {'word':no_punct,'punct':punct}


# CONVERT the cmudict prondict into just numbers
def strip_letters(ls):
    #print "strip_letters"
    nm = ''
    for ws in ls:
        #print "ws",ws
        for ch in list(ws):
            #print "ch",ch
            if ch.isdigit():
                nm=nm+ch
                #print "ad to nm",nm, type(nm)
    return nm


# TESTING  results 
# i do not correct for the '2'
line = "This day (the year I dare not tell)"
print parseStressOfLine(line)
line = "Apollo play'd the midwife's part;"
print parseStressOfLine(line)
line = "Into the world Corinna fell,"
print parseStressOfLine(line)


""" 

OUTPUT 

This day (the year I dare not tell)
{'stress': '01***(*011111***)*', 'stress_no_punct': '01011111'}
Apollo play'd the midwife's part;
{'stress': "0101*'d*01211***;*", 'stress_no_punct': '010101211'}
Into the world Corinna fell,
{'stress': '01012101*,*', 'stress_no_punct': '01012101'}