def symspell_train(self, words_counter: Dict[str, int]) -> 'SymSpell':
"""
Train symspell model
Args:
words_counter (dict): dict of word and their count
Returns:
(SymSpell)
"""
self.symspell_model = SymSpell(config=self.config, script=self.language)
self.symspell_model.create_dictionary_from_words(words_counter)
return self.symspell_model
def __init__(self, language: str = 'en') -> None:
self.language = language
if language not in PhonemeModel.supported_languages + ['en']:
raise Exception(f'"{language}" language is not yet supported')
self.config = None
self.set_default_config()
self.symspell_model = SymSpell(config=self.config,
script=self.language)
self.phoneme_model = None
if self.language in PhonemeModel.supported_languages:
self.phoneme_model = PhonemeModel(config=self.config,
script=self.language)
self.context_model = None
class SpellCorrectionModel(object):
"""
Model to find and correct miss-spelled words utter in text from two words corpus, domain and global dictionary.
For each word in text which do not belong to either of domain or global dictionary, model considered it as
miss-spelled word and try to find out most suitable suggestion first from domain dict, and then from global.
Current model works on 4 key algorithms, which are
1) Symspell Model - Suggest words based on simple edit distance.
2) Phoneme Model - Suggest words based on Soundex code of word.
3) Context Model - Model to suggest most suitable word from suggestion list given by above two model
4) Normalize Model - Model to Normalize unigrams which also occurs as bigrams or trigrams to split form.
for example - Login to log in.
"""
def __init__(self, language: str = 'en') -> None:
self.language = language
if language not in PhonemeModel.supported_languages + ['en']:
raise Exception(f'"{language}" language is not yet supported')
self.config = None
self.set_default_config()
self.symspell_model = SymSpell(config=self.config,
script=self.language)
self.phoneme_model = None
if self.language in PhonemeModel.supported_languages:
self.phoneme_model = PhonemeModel(config=self.config,
script=self.language)
self.context_model = None
def set_default_config(self):
self.config = Config
def symspell_train(self, words_counter: Dict[str, int]) -> 'SymSpell':
"""
Train symspell model
Args:
words_counter (dict): dict of word and their count
Returns:
(SymSpell)
"""
self.symspell_model = SymSpell(config=self.config,
script=self.language)
self.symspell_model.create_dictionary_from_words(words_counter)
return self.symspell_model
def phoneme_train(self, words_counter: Dict[str, int]) -> 'PhonemeModel':
"""
Train phoneme model
Args:
words_counter (dict): dict of word and their count
Returns:
(PhonemeModel)
"""
if self.language not in PhonemeModel.supported_languages:
return self.phoneme_model
self.phoneme_model = PhonemeModel(config=self.config,
script=self.language)
self.phoneme_model.create_phoneme_dictionary_from_words(words_counter)
return self.phoneme_model
def context_train(self, texts: List[str]) -> 'ContextModel':
"""
Train context model
Args:
texts (list): list of text
Returns:
(ContextModel)
"""
self.context_model = ContextModel()
self.context_model.create_model_dict(texts)
return self.context_model
def context_suggestion(
self, text: str,
suggestions_dict: Dict[str,
List[str]]) -> Tuple[str, Dict[str, str]]:
"""
Get context suggested text for given text and suggestions dict with misspelled words and their list
of suggestions
Args:
text (str): text
suggestions_dict: dict with misspelled words and their list of suggestions
Returns:
(str): contextually corrected text
(dict): token level suggestion from context model
Examples:
>>> text = 'i wnt to pley kriket'
>>> suggestions_dict = {'wnt': ['want', 'went', 'what'], 'pley': ['play'], 'kriket': ['cricket']}
>>> SpellCorrectionModel().context_suggestion(text, suggestions_dict)
(i want to play cricket, {'wnt': 'want', 'pley: 'play', 'kriket': 'cricket'})
"""
if not self.context_model or not suggestions_dict:
suggestions_dict = {
key: value[0]
for key, value in suggestions_dict.items()
}
return text, suggestions_dict
text, context_corrections = self.context_model.context_spell_correct(
text, suggestions_dict)
return text, context_corrections
def train(self, data: Union[List[str], Dict[str, int]], **kwargs):
"""
Train all models of spellcorrection
Args:
data (list|dict): list of text or dict having word and their count
Returns:
None
"""
logger.debug("Spello training started..")
words_counter = {}
if not isinstance(data, list) and not isinstance(data, dict):
raise ValueError(
'Argument `data` should be either List[str] or Dict[str, int]')
if isinstance(data, list):
texts = [get_clean_text(text) for text in data]
lower_case_texts = [text.lower() for text in texts]
logger.debug("Context model training started ...")
# Context model get trained only when list of text are given for training
# train context model: find most probable correct word for given suggestions for each word in texts
# based on context word
self.context_train(lower_case_texts)
tokens = " ".join(lower_case_texts).strip().split()
words_counter = dict(Counter(tokens))
elif isinstance(data, dict):
words_counter = {
word.lower(): count
for word, count in data.items()
}
words_counter = {
word: count
for word, count in words_counter.items()
if self.config.min_length_for_spellcorrection
}
logger.debug("Symspell training started ...")
# train symspell model: give suggestion based on edit distance
self.symspell_train(words_counter)
logger.debug("Phoneme training started ...")
# train phoneme model: give suggestion for similar sounding words
self.phoneme_train(words_counter)
logger.debug("Spello training completed successfully ...")
def _correct_word(self, word):
"""
Suggest words for given word, follow below steps:
- check if length of word is eligible for correction from min max length allowed from config,
else return None
- if word belong to domain word correction (created for words which occur less in domain words and are
replaced with most similar words, either from domain or global), return suggested word from dict
- if word exists in global dictionary, return None
- if word exists in domain dictionary, return None
- get suggestions from domain symspell & phoneme
- if not suggestions from domain, get suggestions from global symspell and phoneme
- filter top 5 suggestions
- also add top 3 suggestions from global data up to min edit distance of above dict
- return final suggestions
Args:
word (str): word to be corrected
Returns:
(list): list of suggested words
"""
if ((self.config.min_length_for_spellcorrection > len(word))
or (len(word) > self.config.max_length_for_spellcorrection)
or (re.search(r'\d', word))):
return None
phoneme = self.phoneme_model.spell_correct(word)
if phoneme.is_correct:
return None
phoneme_suggestions = phoneme.suggestions
symspell = self.symspell_model.spell_correct(word)
if symspell.is_correct:
return None
symspell_suggestions = symspell.suggestions
logger.debug(f"Symspell suggestions: {symspell_suggestions}")
logger.debug(f"Phoneme suggestions: {phoneme_suggestions}")
phoneme_suggestions = [(word, edit_distance, 0)
for (word, edit_distance) in phoneme_suggestions
]
symspell_suggestions = [(word, edit_distance, 1)
for (word,
edit_distance) in symspell_suggestions]
suggestions = list(phoneme_suggestions + symspell_suggestions)
suggestions.sort(key=operator.itemgetter(1, 2))
# filtering duplicate suggestions
unique_suggestions = []
for suggestion in suggestions:
if suggestion[0] not in unique_suggestions:
unique_suggestions.append(suggestion[0])
suggestions = unique_suggestions
return suggestions
def spell_correct(self, text: str, verbose=0) -> Dict[str, Any]:
"""
Get spell corrected text, and dict of token level suggestion
Args:
text (str):
verbose (int): define verbose level
Returns:
(str): spell corrected text
(dict): dict of token level suggestion
Examples:
>>> text = 'i wnt to play kricket'
>>> SpellCorrectionModel().spell_correct(text)
{
'original_text': 'i wnt to play kricket',
'corrected_text': 'i want to play cricket',
'correction_dict': {'wnt': 'want', 'kricket': 'cricket'}
}
"""
levels = [logging.CRITICAL, logging.ERROR, logging.INFO, logging.DEBUG]
verbosity = min(verbose, len(levels) - 1)
with loglevel(levels[verbosity]):
spellcorrection_result = {
ORIGINAL_TEXT: text,
CORRECTED_TEXT: text,
CORRECTIONS_DICT: {}
}
suggestions_dict = {}
# clean_text = get_clean_text(text)
clean_text = text
tokens = clean_text.split()
for token in tokens:
lowercase_token = token.lower()
token_suggestion = self._correct_word(lowercase_token)
if token_suggestion:
suggestions_dict[lowercase_token] = token_suggestion
logger.debug(
f"Suggestions dict from phoneme and symspell are: {suggestions_dict}"
)
context_corrected_text, context_corrections = self.context_suggestion(
clean_text, suggestions_dict)
logger.debug(f"text after context model: {context_corrected_text}")
spellcorrection_result[CORRECTED_TEXT] = context_corrected_text
spellcorrection_result[CORRECTIONS_DICT] = context_corrections
logger.debug(f"Spell-correction Results {spellcorrection_result}")
return spellcorrection_result
def get_state(self) -> Dict[str, Any]:
return {
'language': self.language,
'symspell_model': self.symspell_model,
'phoneme_model': self.phoneme_model,
'context_model': self.context_model,
'config': self.config
}
def set_state(self, state_dict: Dict[str, Any]) -> None:
self.language = state_dict['language']
self.config = state_dict['config']
self.symspell_model = state_dict['symspell_model']
self.phoneme_model = state_dict['phoneme_model']
self.context_model = state_dict['context_model']
def load(self, model_path: str, **kwargs) -> 'SpellCorrectionModel':
model_state = pickle.load(open(model_path, 'rb'))
self.set_state(model_state)
return self
def save(self, model_save_dir: str, **kwargs) -> str:
save_dir = Path(model_save_dir)
mkdirs(save_dir, err_if_already_exists=False)
state = self.get_state()
pickle.dump(state, open(save_dir / 'model.pkl', "wb"))
path = os.path.join(save_dir, 'model.pkl')
return path