def initialize(self):
self._vocabulary_sizes_for_ngram_item_types = self._find_vocabulary_sizes(self._ngram_item_types)
if logger.isEnabledFor(logging.DEBUG):
logger.debug("Found vocabulary sizes for ngram types : " + str(self._vocabulary_sizes_for_ngram_item_types))
for context_type in self._ngram_item_types:
for context_is_leading in (True, False):
for target_type in self._ngram_item_types:
ngram_type, type_key = self._get_ngram_type_and_key(context_is_leading, context_type, target_type)
if type_key in self._smoothers_for_ngram_types:
# stuff already calculated, smoother already created!
continue
distinct_ngram_count_for_ngram_type = NgramTypeFrequencyFinder.find_distinct_count(
self._collection, ngram_type
)
possible_ngram_count_for_ngram_type = reduce(
operator.mul,
[
self._vocabulary_sizes_for_ngram_item_types[ngram_type_item]
for ngram_type_item in ngram_type
],
)
frequency_of_frequency_0 = possible_ngram_count_for_ngram_type - distinct_ngram_count_for_ngram_type
logger.debug(" Distinct ngram count for ngram type = " + str(distinct_ngram_count_for_ngram_type))
logger.debug(" Possible ngram count for ngram type = " + str(possible_ngram_count_for_ngram_type))
logger.debug(" Frequency of frequency 0 (unseen) = " + str(frequency_of_frequency_0))
frequencies_of_frequencies_for_ngram_type = {}
for i in range(1, self._smoothing_threshold + 2):
frequencies_of_frequencies_for_ngram_type[i] = self._find_frequency_of_frequency(ngram_type, i)
smoother = SimpleGoodTuringSmoother(
self._smoothing_threshold, frequencies_of_frequencies_for_ngram_type, frequency_of_frequency_0
)
self._smoothers_for_ngram_types[type_key] = smoother
for ngram_type_key, smoother in self._smoothers_for_ngram_types.iteritems():
smoother.initialize(PLOTTING_MODE)
if logger.isEnabledFor(logging.DEBUG):
for ngram_type_key, smoother in self._smoothers_for_ngram_types.iteritems():
# convert default dict to normal dict and then pprint it
logger.debug(
"Found frequencies of ngram frequencies for {}: "
+ pprint.pformat(json.loads(json.dumps(smoother._frequencies_of_frequencies)))
)
logger.debug("Found unseen for {}: {}".format(smoother._unseen_count))
if logger.isEnabledFor(logging.DEBUG):
for ngram_type_key, smoother in self._smoothers_for_ngram_types.iteritems():
# convert default dict to normal dict and then pprint it
logger.debug(
"Loglin regression coefficient m for {}: ".format(ngram_type_key, smoother._loglinregression_m)
)
logger.debug(
"Loglin regression coefficient c for {}: ".format(ngram_type_key, smoother._loglinregression_c)
)
def test_with_larger_values_sc_2(self):
smoother = SimpleGoodTuringSmoother(K, {
1: 16181,
2: 2213,
3: 870,
4: 431,
5: 304,
6: 202
}, 2111251811)
smoother.initialize()
for i in range(0, K + 5):
logger.info("c_{} : {}, \t c*_{} : {}".format(
i, i, i, smoother.smooth(i)))
def test_with_small_values(self):
smoother = SimpleGoodTuringSmoother(K, {
1: 10,
2: 5,
3: 3,
4: 2,
5: 1,
6: 0
}, 100)
smoother.initialize()
for i in range(0, K + 5):
logger.info("c_{} : {}, \t c*_{} : {}".format(
i, i, i, smoother.smooth(i)))
def test_with_larger_values(self):
smoother = SimpleGoodTuringSmoother(K, {
1: 268,
2: 112,
3: 70,
4: 41,
5: 24,
6: 14,
7: 15,
400: 1,
1918: 1
}, 1000)
smoother.initialize()
for i in range(0, K + 5):
logger.info("c_{} : {}, \t c*_{} : {}".format(
i, i, i, smoother.smooth(i)))
def test_with_zero_frequencies_in_between(self):
smoother = SimpleGoodTuringSmoother(K, {
1: 268,
2: 0,
3: 70,
4: 0,
5: 24,
6: 14,
7: 15,
400: 1,
1918: 1
}, 1000)
smoother.initialize()
for i in range(0, K + 5):
logger.info("c_{} : {}, \t c*_{} : {}".format(
i, i, i, smoother.smooth(i)))
def initialize(self):
self._vocabulary_sizes_for_ngram_item_types = self._find_vocabulary_sizes(
self._ngram_item_types)
if logger.isEnabledFor(logging.DEBUG):
logger.debug("Found vocabulary sizes for ngram types : " +
str(self._vocabulary_sizes_for_ngram_item_types))
for context_type in self._ngram_item_types:
for context_is_leading in (True, False):
for target_type in self._ngram_item_types:
ngram_type, type_key = self._get_ngram_type_and_key(
context_is_leading, context_type, target_type)
if type_key in self._smoothers_for_ngram_types:
# stuff already calculated, smoother already created!
continue
distinct_ngram_count_for_ngram_type = NgramTypeFrequencyFinder.find_distinct_count(
self._collection, ngram_type)
possible_ngram_count_for_ngram_type = reduce(
operator.mul, [
self._vocabulary_sizes_for_ngram_item_types[
ngram_type_item]
for ngram_type_item in ngram_type
])
frequency_of_frequency_0 = possible_ngram_count_for_ngram_type - distinct_ngram_count_for_ngram_type
logger.debug(" Distinct ngram count for ngram type = " +
str(distinct_ngram_count_for_ngram_type))
logger.debug(" Possible ngram count for ngram type = " +
str(possible_ngram_count_for_ngram_type))
logger.debug(" Frequency of frequency 0 (unseen) = " +
str(frequency_of_frequency_0))
frequencies_of_frequencies_for_ngram_type = {}
for i in range(1, self._smoothing_threshold + 2):
frequencies_of_frequencies_for_ngram_type[
i] = self._find_frequency_of_frequency(
ngram_type, i)
smoother = SimpleGoodTuringSmoother(
self._smoothing_threshold,
frequencies_of_frequencies_for_ngram_type,
frequency_of_frequency_0)
self._smoothers_for_ngram_types[type_key] = smoother
for ngram_type_key, smoother in self._smoothers_for_ngram_types.iteritems(
):
smoother.initialize(PLOTTING_MODE)
if logger.isEnabledFor(logging.DEBUG):
for ngram_type_key, smoother in self._smoothers_for_ngram_types.iteritems(
):
# convert default dict to normal dict and then pprint it
logger.debug(
"Found frequencies of ngram frequencies for {}: " +
pprint.pformat(
json.loads(
json.dumps(smoother._frequencies_of_frequencies))))
logger.debug("Found unseen for {}: {}".format(
smoother._unseen_count))
if logger.isEnabledFor(logging.DEBUG):
for ngram_type_key, smoother in self._smoothers_for_ngram_types.iteritems(
):
# convert default dict to normal dict and then pprint it
logger.debug("Loglin regression coefficient m for {}: ".format(
ngram_type_key, smoother._loglinregression_m))
logger.debug("Loglin regression coefficient c for {}: ".format(
ngram_type_key, smoother._loglinregression_c))
class SimpleGoodTuringContextlessDistributionSmoother(
ContextlessDistributionSmoother):
AVG_PARSE_RESULTS_FOR_A_WORD = 6 # avg parse result count for a word
AVG_WORDS_FOR_A_LEXEME = 50 # avg word count for a lexeme
AVG_WORDS_FOR_A_STEM = 10 # avg word count for a stem
def __init__(self, smoothing_threshold, unigram_collection):
self._smoothing_threshold = smoothing_threshold
self._unigram_collection = unigram_collection
assert self._smoothing_threshold and self._smoothing_threshold > 1
def initialize(self):
logger.debug(
"Initializing SimpleGoodTuringContextlessDistributionSmoother for K:{}, AVG_PARSE_RESULTS_FOR_A_WORD:{}, AVG_WORDS_FOR_A_LEXEME:{}"
.format(self._smoothing_threshold,
self.AVG_PARSE_RESULTS_FOR_A_WORD,
self.AVG_WORDS_FOR_A_LEXEME))
distinct_parse_result_count = NgramTypeFrequencyFinder.find_distinct_parse_result_count(
self._unigram_collection)
distinct_word_count = NgramTypeFrequencyFinder.find_distinct_word_count(
self._unigram_collection)
distinct_lexeme_count = NgramTypeFrequencyFinder.find_distinct_count(
self._unigram_collection, ['lemma_root'])
distinct_stem_count = NgramTypeFrequencyFinder.find_distinct_count(
self._unigram_collection, ['stem'])
possible_word_count_estimate_from_lexemes = distinct_lexeme_count * self.AVG_WORDS_FOR_A_LEXEME
possible_word_count_estimate_from_stems = distinct_stem_count * self.AVG_WORDS_FOR_A_STEM
possible_word_count_estimate = possible_word_count_estimate_from_stems + possible_word_count_estimate_from_lexemes
unseen_word_count = possible_word_count_estimate - distinct_word_count
possible_parse_result_count_estimate = possible_word_count_estimate * self.AVG_PARSE_RESULTS_FOR_A_WORD
unseen_parse_result_count = possible_parse_result_count_estimate - distinct_parse_result_count
logger.debug("Found {} distinct parse results".format(
distinct_parse_result_count))
logger.debug("Found {} distinct words".format(distinct_word_count))
logger.debug("Estimated possible parse result count : {}".format(
possible_parse_result_count_estimate))
logger.debug("Estimated unseen parse result count : {}".format(
unseen_parse_result_count))
logger.debug("Found {} distinct lexemes".format(distinct_lexeme_count))
logger.debug("Estimated possible word count from lexemes: {}".format(
possible_word_count_estimate_from_lexemes))
logger.debug("Estimated possible word count from stems: {}".format(
possible_word_count_estimate_from_stems))
logger.debug("Estimated possible word count: {}".format(
possible_word_count_estimate))
logger.debug(
"Estimated unseen word count : {}".format(unseen_word_count))
frequencies_of_parse_result_frequencies = {
1: distinct_parse_result_count
}
frequencies_of_word_frequencies = {1: distinct_word_count}
for i in range(2, self._smoothing_threshold + 2):
frequencies_of_parse_result_frequencies[
i] = NgramTypeFrequencyFinder.find_frequency_of_parse_result_frequency(
self._unigram_collection, i)
frequencies_of_word_frequencies[
i] = NgramTypeFrequencyFinder.find_frequency_of_word_frequency(
self._unigram_collection, i)
logger.debug("Frequencies of parse result frequencies")
logger.debug(pformat(frequencies_of_parse_result_frequencies))
logger.debug("Frequencies of word frequencies")
logger.debug(pformat(frequencies_of_word_frequencies))
self._parse_result_count_smoother = SimpleGoodTuringSmoother(
self._smoothing_threshold, frequencies_of_parse_result_frequencies,
unseen_parse_result_count)
self._word_count_smoother = SimpleGoodTuringSmoother(
self._smoothing_threshold, frequencies_of_word_frequencies,
unseen_word_count)
self._parse_result_count_smoother.initialize()
self._word_count_smoother.initialize()
def smooth_parse_result_occurrence_count(self,
parse_result_occurrence_count):
if parse_result_occurrence_count > self._smoothing_threshold:
return parse_result_occurrence_count
return self._parse_result_count_smoother.smooth(
parse_result_occurrence_count)
def smooth_word_occurrence_count(self, word_occurrence_count):
if word_occurrence_count > self._smoothing_threshold:
return word_occurrence_count
return self._word_count_smoother.smooth(word_occurrence_count)
def test_with_zero_frequencies_in_between(self):
smoother = SimpleGoodTuringSmoother(K, {1: 268, 2: 0, 3: 70, 4: 0, 5: 24, 6: 14, 7: 15, 400: 1, 1918: 1}, 1000)
smoother.initialize()
for i in range(0, K + 5):
logger.info("c_{} : {}, \t c*_{} : {}".format(i, i, i, smoother.smooth(i)))
def test_with_larger_values_sc_2(self):
smoother = SimpleGoodTuringSmoother(K, {1: 16181, 2: 2213, 3: 870, 4: 431, 5: 304, 6: 202}, 2111251811)
smoother.initialize()
for i in range(0, K + 5):
logger.info("c_{} : {}, \t c*_{} : {}".format(i, i, i, smoother.smooth(i)))
def test_with_larger_values(self):
smoother = SimpleGoodTuringSmoother(K, {1: 268, 2: 112, 3: 70, 4: 41, 5: 24, 6: 14, 7: 15, 400: 1, 1918: 1}, 1000)
smoother.initialize()
for i in range(0, K + 5):
logger.info("c_{} : {}, \t c*_{} : {}".format(i, i, i, smoother.smooth(i)))
def test_with_small_values(self):
smoother = SimpleGoodTuringSmoother(K, {1: 10, 2: 5, 3: 3, 4: 2, 5: 1, 6: 0}, 100)
smoother.initialize()
for i in range(0, K + 5):
logger.info("c_{} : {}, \t c*_{} : {}".format(i, i, i, smoother.smooth(i)))
class SimpleGoodTuringContextlessDistributionSmoother(ContextlessDistributionSmoother):
AVG_PARSE_RESULTS_FOR_A_WORD = 6 # avg parse result count for a word
AVG_WORDS_FOR_A_LEXEME = 50 # avg word count for a lexeme
AVG_WORDS_FOR_A_STEM = 10 # avg word count for a stem
def __init__(self, smoothing_threshold, unigram_collection):
self._smoothing_threshold = smoothing_threshold
self._unigram_collection = unigram_collection
assert self._smoothing_threshold and self._smoothing_threshold > 1
def initialize(self):
logger.debug(
"Initializing SimpleGoodTuringContextlessDistributionSmoother for K:{}, AVG_PARSE_RESULTS_FOR_A_WORD:{}, AVG_WORDS_FOR_A_LEXEME:{}".format(
self._smoothing_threshold, self.AVG_PARSE_RESULTS_FOR_A_WORD, self.AVG_WORDS_FOR_A_LEXEME))
distinct_parse_result_count = NgramTypeFrequencyFinder.find_distinct_parse_result_count(
self._unigram_collection)
distinct_word_count = NgramTypeFrequencyFinder.find_distinct_word_count(self._unigram_collection)
distinct_lexeme_count = NgramTypeFrequencyFinder.find_distinct_count(self._unigram_collection, ['lemma_root'])
distinct_stem_count = NgramTypeFrequencyFinder.find_distinct_count(self._unigram_collection, ['stem'])
possible_word_count_estimate_from_lexemes = distinct_lexeme_count * self.AVG_WORDS_FOR_A_LEXEME
possible_word_count_estimate_from_stems = distinct_stem_count * self.AVG_WORDS_FOR_A_STEM
possible_word_count_estimate = possible_word_count_estimate_from_stems + possible_word_count_estimate_from_lexemes
unseen_word_count = possible_word_count_estimate - distinct_word_count
possible_parse_result_count_estimate = possible_word_count_estimate * self.AVG_PARSE_RESULTS_FOR_A_WORD
unseen_parse_result_count = possible_parse_result_count_estimate - distinct_parse_result_count
logger.debug("Found {} distinct parse results".format(distinct_parse_result_count))
logger.debug("Found {} distinct words".format(distinct_word_count))
logger.debug("Estimated possible parse result count : {}".format(possible_parse_result_count_estimate))
logger.debug("Estimated unseen parse result count : {}".format(unseen_parse_result_count))
logger.debug("Found {} distinct lexemes".format(distinct_lexeme_count))
logger.debug("Estimated possible word count from lexemes: {}".format(possible_word_count_estimate_from_lexemes))
logger.debug("Estimated possible word count from stems: {}".format(possible_word_count_estimate_from_stems))
logger.debug("Estimated possible word count: {}".format(possible_word_count_estimate))
logger.debug("Estimated unseen word count : {}".format(unseen_word_count))
frequencies_of_parse_result_frequencies = {1: distinct_parse_result_count}
frequencies_of_word_frequencies = {1: distinct_word_count}
for i in range(2, self._smoothing_threshold + 2):
frequencies_of_parse_result_frequencies[
i] = NgramTypeFrequencyFinder.find_frequency_of_parse_result_frequency(self._unigram_collection, i)
frequencies_of_word_frequencies[i] = NgramTypeFrequencyFinder.find_frequency_of_word_frequency(
self._unigram_collection, i)
logger.debug("Frequencies of parse result frequencies")
logger.debug(pformat(frequencies_of_parse_result_frequencies))
logger.debug("Frequencies of word frequencies")
logger.debug(pformat(frequencies_of_word_frequencies))
self._parse_result_count_smoother = SimpleGoodTuringSmoother(self._smoothing_threshold,
frequencies_of_parse_result_frequencies,
unseen_parse_result_count)
self._word_count_smoother = SimpleGoodTuringSmoother(self._smoothing_threshold, frequencies_of_word_frequencies,
unseen_word_count)
self._parse_result_count_smoother.initialize()
self._word_count_smoother.initialize()
def smooth_parse_result_occurrence_count(self, parse_result_occurrence_count):
if parse_result_occurrence_count > self._smoothing_threshold:
return parse_result_occurrence_count
return self._parse_result_count_smoother.smooth(parse_result_occurrence_count)
def smooth_word_occurrence_count(self, word_occurrence_count):
if word_occurrence_count > self._smoothing_threshold:
return word_occurrence_count
return self._word_count_smoother.smooth(word_occurrence_count)