Merge pull request #79 from QData/wordnet-transformation

Wordnet transformation

scripts/run_attack_args_helper.py

@@ -74,6 +74,7 @@
 }
 
 TRANSFORMATION_CLASS_NAMES = {
+    'word-swap-wordnet':               'textattack.transformations.WordSwapWordNet',
     'word-swap-embedding':             'textattack.transformations.WordSwapEmbedding',
     'word-swap-homoglyph':             'textattack.transformations.WordSwapHomoglyph',
     'word-swap-neighboring-char-swap': 'textattack.transformations.WordSwapNeighboringCharacterSwap',
@@ -283,6 +284,7 @@ def parse_logger_from_args(args):# Create logger
         attack_logger.add_output_csv(csv_path, color_method)
         print('Logging to CSV at path {}.'.format(csv_path))
 
+
     # Visdom
     if args.enable_visdom:
         attack_logger.enable_visdom()

textattack/transformations/__init__.py

@@ -5,4 +5,5 @@
 from .word_swap_neighboring_character_swap import WordSwapNeighboringCharacterSwap
 from .word_swap_random_character_deletion import WordSwapRandomCharacterDeletion
 from .word_swap_random_character_insertion import WordSwapRandomCharacterInsertion
-from .word_swap_random_character_substitution import WordSwapRandomCharacterSubstitution
+from .word_swap_random_character_substitution import WordSwapRandomCharacterSubstitution
+from .word_swap_wordnet import WordSwapWordNet

textattack/transformations/word_swap.py

@@ -50,16 +50,26 @@ def __call__(self, tokenized_text, indices_to_replace=None):
         word_swaps = []
         for i in indices_to_replace:
             word_to_replace = words[i]
+            # Don't replace stopwords.
             if not self.replace_stopwords and word_to_replace.lower() in self.stopwords:
                 continue
             replacement_words = self._get_replacement_words(word_to_replace)
             new_tokenized_texts = []
             for r in replacement_words:
+                # Don't replace with numbers, punctuation, or other non-letter characters.
+                if not is_word(r):
+                    continue
                 new_tokenized_texts.append(tokenized_text.replace_word_at_index(i, r))
             transformations.extend(new_tokenized_texts)
 
         return transformations
 
 
     def extra_repr_keys(self): 
-        return ['replace_stopwords']
+        return ['replace_stopwords']
+
+def is_word(s):
+    """ String `s` counts as a word if it has at least one letter. """
+    for c in s:
+        if c.isalpha(): return True
+    return False 

textattack/transformations/word_swap_homoglyph.py

@@ -1,11 +1,10 @@
 import numpy as np
-import os
 
 from textattack.shared import utils
 from textattack.transformations.word_swap import WordSwap
 
 class WordSwapHomoglyph(WordSwap):
-    """ Transforms an input by replacing its words with visually-similar words using homoglyph swaps.
+    """ Transforms an input by replacing its words with visually similar words using homoglyph swaps.
     """
 
     def __init__(self, replace_stopwords=False):

textattack/transformations/word_swap_neighboring_character_swap.py

@@ -1,5 +1,4 @@
 import numpy as np
-import os
 
 from textattack.shared import utils
 from textattack.transformations.word_swap import WordSwap
@@ -36,4 +35,4 @@ def _get_replacement_words(self, word):
         return candidate_words
 
     def extra_repr_keys(self): 
-        return super().extra_repr_keys() + ['random_one']
+        return super().extra_repr_keys() + ['random_one']

textattack/transformations/word_swap_random_character_deletion.py

@@ -1,5 +1,4 @@
 import numpy as np
-import os
 
 from textattack.shared import utils
 from textattack.transformations.word_swap import WordSwap
@@ -35,4 +34,4 @@ def _get_replacement_words(self, word):
         return candidate_words
 
     def extra_repr_keys(self): 
-        return super().extra_repr_keys() + ['random_one']
+        return super().extra_repr_keys() + ['random_one']

textattack/transformations/word_swap_random_character_insertion.py

@@ -1,5 +1,4 @@
 import numpy as np
-import os
 
 from textattack.shared import utils
 from textattack.transformations.word_swap import WordSwap
@@ -32,4 +31,4 @@ def _get_replacement_words(self, word):
                 candidate_word = word[:i] + self._get_random_letter() + word[i:]
                 candidate_words.append(candidate_word)
 
-        return candidate_words
+        return candidate_words

textattack/transformations/word_swap_random_character_substitution.py

@@ -1,5 +1,4 @@
 import numpy as np
-import os
 
 from textattack.shared import utils
 from textattack.transformations.word_swap import WordSwap
@@ -33,4 +32,4 @@ def _get_replacement_words(self, word):
                 candidate_word = word[:i] + self._get_random_letter() + word[i+1:]
                 candidate_words.append(candidate_word)
 
-        return candidate_words
+        return candidate_words

textattack/transformations/word_swap_wordnet.py

@@ -0,0 +1,15 @@
+from nltk.corpus import wordnet
+from textattack.transformations.word_swap import WordSwap
+
+class WordSwapWordNet(WordSwap):
+    """ Transforms an input by replacing its words with synonyms provided by WordNet.
+    """
+
+    def _get_replacement_words(self, word, random=False):
+        """ Returns a list containing all possible words with 1 character replaced by a homoglyph.
+        """
+        synonyms = set()
+        for syn in wordnet.synsets(word): 
+            for l in syn.lemmas(): 
+                synonyms.add(l.name()) 
+        return list(synonyms)