Newer
Older
Donato Meoli
a validé
Statistical Language Processing tools (Chapter 22)
We define Unigram and Ngram text models, use them to generate random text,
and show the Viterbi algorithm for segmentation of letters into words.
Then we show a very simple Information Retrieval system, and an example
working on a tiny sample of Unix manual pages.
"""
Donato Meoli
a validé
import re
from collections import defaultdict
import numpy as np
import search
from probabilistic_learning import CountingProbDist
from utils import hashabledict
"""This is a discrete probability distribution over words, so you
can add, sample, or get P[word], just like with CountingProbDist. You can
also generate a random text, n words long, with P.samples(n)."""
def __init__(self, observations, default=0):
# Call CountingProbDist constructor,
# passing the observations and default parameters.
super(UnigramWordModel, self).__init__(observations, default)
def samples(self, n):
"""Return a string of n words, random according to the model."""
return ' '.join(self.sample() for i in range(n))
"""This is a discrete probability distribution over n-tuples of words.
You can add, sample or get P[(word1, ..., wordn)]. The method P.samples(n)
builds up an n-word sequence; P.add_cond_prob and P.add_sequence add data."""
def __init__(self, n, observation_sequence=None, default=0):
# In addition to the dictionary of n-tuples, cond_prob is a
# mapping from (w1, ..., wn-1) to P(wn | w1, ... wn-1)
CountingProbDist.__init__(self, default=default)
self.n = n
self.add_sequence(observation_sequence or [])
# __getitem__, top, sample inherited from CountingProbDist
# Note that they deal with tuples, not strings, as inputs
"""Build the conditional probabilities P(wn | (w1, ..., wn-1)"""
if ngram[:-1] not in self.cond_prob:
self.cond_prob[ngram[:-1]] = CountingProbDist()
self.cond_prob[ngram[:-1]].add(ngram[-1])
def add_sequence(self, words):
"""Add each tuple words[i:i+n], using a sliding window."""
n = self.n
t = tuple(words[i:i + n])
self.add(t)
self.add_cond_prob(t)
def samples(self, nwords):
"""Generate an n-word sentence by picking random samples
according to the model. At first pick a random n-gram and
from then on keep picking a character according to
P(c|wl-1, wl-2, ..., wl-n+1) where wl-1 ... wl-n+1 are the
last n - 1 words in the generated sentence so far."""
n = self.n
output = list(self.sample())
for i in range(n, nwords):
next_word = self.cond_prob[tuple(last)].sample()
output.append(next_word)
return ' '.join(output)
"""Add an empty space to every word to catch the beginning of words."""
def __init__(self, observation_sequence=None, default=0):
CountingProbDist.__init__(self, default=default)
self.n = 1
self.cond_prob = defaultdict()
self.add_sequence(observation_sequence or [])
for word in words:
for char in word:
self.add(char)
# ______________________________________________________________________________
def viterbi_segment(text, P):
"""Find the best segmentation of the string of characters, given the
# best[i] = best probability for text[0:i]
# words[i] = best word ending at position i
n = len(text)
words = [''] + list(text)
best = [1.0] + [0.0] * n
# Fill in the vectors best words via dynamic programming
for j in range(0, i):
w = text[j:i]
curr_score = P[w] * best[i - len(w)]
if curr_score >= best[i]:
best[i] = curr_score
words[i] = w
# Now recover the sequence of best words
sequence = []
while i > 0:
sequence[0:0] = [words[i]]
i = i - len(words[i])
# Return sequence of best words and overall probability
return sequence, best[-1]
# ______________________________________________________________________________
class IRSystem:
"""A very simple Information Retrieval System, as discussed in Sect. 23.2.
The constructor s = IRSystem('the a') builds an empty system with two
stopwords. Next, index several documents with s.index_document(text, url).
Then ask queries with s.query('query words', n) to retrieve the top n
matching documents. Queries are literal words from the document,
except that stopwords are ignored, and there is one special syntax:
The query "learn: man cat", for example, runs "man cat" and indexes it."""
def __init__(self, stopwords='the a of'):
"""Create an IR System. Optionally specify stopwords."""
# index is a map of {word: {docid: count}}, where docid is an int,
self.index = defaultdict(lambda: defaultdict(int))
self.stopwords = set(words(stopwords))
self.documents = []
def index_collection(self, filenames):
prefix = os.path.dirname(__file__)
for filename in filenames:
self.index_document(open(filename).read(), os.path.relpath(filename, prefix))
def index_document(self, text, url):
title = text[:text.index('\n')].strip()
docwords = words(text)
self.documents.append(Document(title, url, len(docwords)))
for word in docwords:
if word not in self.stopwords:
def query(self, query_text, n=10):
"""Return a list of n (score, docid) pairs for the best matches.
Also handle the special syntax for 'learn: command'."""
if query_text.startswith("learn:"):
doctext = os.popen(query_text[len("learn:"):], 'r').read()
self.index_document(doctext, query_text)
return []
qwords = [w for w in words(query_text) if w not in self.stopwords]
shortest = min(qwords, key=lambda w: len(self.index[w]))
docids = self.index[shortest]
return heapq.nlargest(n, ((self.total_score(qwords, docid), docid) for docid in docids))
"""Compute a score for this word on the document with this docid."""
# There are many options; here we take a very simple approach
Donato Meoli
a validé
return np.log(1 + self.index[word][docid]) / np.log(1 + self.documents[docid].nwords)
"""Compute the sum of the scores of these words on the document with this docid."""
return sum(self.score(word, docid) for word in words)
def present(self, results):
for (score, docid) in results:
doc = self.documents[docid]
print("{:5.2}|{:25} | {}".format(100 * score, doc.url, doc.title[:45].expandtabs()))
def present_results(self, query_text, n=10):
"""Get results for the query and present them."""
self.present(self.query(query_text, n))
class UnixConsultant(IRSystem):
"""A trivial IR system over a small collection of Unix man pages."""
def __init__(self):
IRSystem.__init__(self, stopwords="how do i the a of")
import os
aima_root = os.path.dirname(__file__)
mandir = os.path.join(aima_root, 'aima-data/MAN/')
man_files = [mandir + f for f in os.listdir(mandir) if f.endswith('.txt')]
self.index_collection(man_files)
class Document:
"""Metadata for a document: title and url; maybe add others later."""
def __init__(self, title, url, nwords):
self.title = title
self.url = url
self.nwords = nwords
def words(text, reg=re.compile('[a-z0-9]+')):
"""Return a list of the words in text, ignoring punctuation and
converting everything to lowercase (to canonicalize).
>>> words("``EGAD!'' Edgar cried.")
['egad', 'edgar', 'cried']
"""
return reg.findall(text.lower())
def canonicalize(text):
"""Return a canonical text: only lowercase letters and blanks.
>>> canonicalize("``EGAD!'' Edgar cried.")
'egad edgar cried'
"""
return ' '.join(words(text))
# ______________________________________________________________________________
# Example application (not in book): decode a cipher.
# A cipher is a code that substitutes one character for another.
# A shift cipher is a rotation of the letters in the alphabet,
# such as the famous rot13, which maps A to N, B to M, etc.
alphabet = 'abcdefghijklmnopqrstuvwxyz'
def shift_encode(plaintext, n):
"""Encode text with a shift cipher that moves each letter up by n letters.
>>> shift_encode('abc z', 1)
'bcd a'
"""
return encode(plaintext, alphabet[n:] + alphabet[:n])
def rot13(plaintext):
"""Encode text by rotating letters by 13 spaces in the alphabet.
>>> rot13('hello')
'uryyb'
>>> rot13(rot13('hello'))
'hello'
"""
return shift_encode(plaintext, 13)
def translate(plaintext, function):
"""Translate chars of a plaintext with the given function."""
result = ""
for char in plaintext:
result += function(char)
return result
def maketrans(from_, to_):
"""Create a translation table and return the proper function."""
trans_table = {}
for n, char in enumerate(from_):
trans_table[char] = to_[n]
return lambda char: trans_table.get(char, char)
def encode(plaintext, code):
"""Encode text using a code which is a permutation of the alphabet."""
trans = maketrans(alphabet + alphabet.upper(), code + code.upper())
def bigrams(text):
"""Return a list of pairs in text (a sequence of letters or words).
>>> bigrams('this')
['th', 'hi', 'is']
>>> bigrams(['this', 'is', 'a', 'test'])
[['this', 'is'], ['is', 'a'], ['a', 'test']]
"""
return [text[i:i + 2] for i in range(len(text) - 1)]
class ShiftDecoder:
"""There are only 26 possible encodings, so we can try all of them,
and return the one with the highest probability, according to a
bigram probability distribution."""
def __init__(self, training_text):
training_text = canonicalize(training_text)
self.P2 = CountingProbDist(bigrams(training_text), default=1)
def score(self, plaintext):
"""Return a score for text based on how common letters pairs are."""
s = 1.0
for bi in bigrams(plaintext):
s = s * self.P2[bi]
return s
def decode(self, ciphertext):
"""Return the shift decoding of text with the best score."""
return max(all_shifts(ciphertext), key=lambda shift: self.score(shift))
def all_shifts(text):
"""Return a list of all 26 possible encodings of text by a shift cipher."""
yield from (shift_encode(text, i) for i, _ in enumerate(alphabet))
class PermutationDecoder:
"""This is a much harder problem than the shift decoder. There are 26!
permutations, so we can't try them all. Instead we have to search.
We want to search well, but there are many things to consider:
Unigram probabilities (E is the most common letter); Bigram probabilities
(TH is the most common bigram); word probabilities (I and A are the most
common one-letter words, etc.); etc.
We could represent a search state as a permutation of the 26 letters,
and alter the solution through hill climbing. With an initial guess
based on unigram probabilities, this would probably fare well. However,
I chose instead to have an incremental representation. A state is
represented as a letter-to-letter map; for example {'z': 'e'} to
represent that 'z' will be translated to 'e'."""
def __init__(self, training_text, ciphertext=None):
self.Pwords = UnigramWordModel(words(training_text))
self.P1 = UnigramWordModel(training_text) # By letter
self.P2 = NgramWordModel(2, words(training_text)) # By letter pair
def decode(self, ciphertext):
"""Search for a decoding of the ciphertext."""
self.ciphertext = canonicalize(ciphertext)
# reduce domain to speed up search
self.chardomain = {c for c in self.ciphertext if c != ' '}
problem = PermutationDecoderProblem(decoder=self)
solution = search.best_first_graph_search(
problem, lambda node: self.score(node.state))
solution.state[' '] = ' '
return translate(self.ciphertext, lambda c: solution.state[c])
"""Score is product of word scores, unigram scores, and bigram scores.
This can get very small, so we use logs and exp."""
# remake code dictionary to contain translation for all characters
full_code = code.copy()
full_code.update({x: x for x in self.chardomain if x not in code})
full_code[' '] = ' '
text = translate(self.ciphertext, lambda c: full_code[c])
# add small positive value to prevent computing log(0)
# TODO: Modify the values to make score more accurate
Donato Meoli
a validé
logP = (sum(np.log(self.Pwords[word] + 1e-20) for word in words(text)) +
sum(np.log(self.P1[c] + 1e-5) for c in text) +
sum(np.log(self.P2[b] + 1e-10) for b in bigrams(text)))
return -np.exp(logP)
class PermutationDecoderProblem(search.Problem):
def __init__(self, initial=None, goal=None, decoder=None):
super().__init__(initial or hashabledict(), goal)
self.decoder = decoder
search_list = [c for c in self.decoder.chardomain if c not in state]
target_list = [c for c in alphabet if c not in state.values()]
# Find the best character to replace
plain_char = max(search_list, key=lambda c: self.decoder.P1[c])
for cipher_char in target_list:
yield (plain_char, cipher_char)
def result(self, state, action):
new_state = hashabledict(state) # copy to prevent hash issues
new_state[action[0]] = action[1]
return new_state
def goal_test(self, state):
"""We're done when all letters in search domain are assigned."""
return len(state) >= len(self.decoder.chardomain)