Negated antonym

requirements.txt

@@ -2,6 +2,7 @@ checklist==0.0.11
 spacy==3.0.0
 numpy
 
+
 # for back_translation
 torch
 

transformations/negated_antonym_perturbation/README.md

@@ -0,0 +1,27 @@
+# Negated Antonym Perturbation 
+This perturbation rephrases the adjectives and adverbs to their negated antonym
+
+Authors: Mayukh Das (Technical University of Braunschweig  / [email protected]) 
+
+## What type of a transformation is this?
+This transformation types of adjectives and adverbs and converts them to their negated antonym.
+Therefore, the transformation retains the semantics of the original text as positives gets converted to negated negatives and negatives gets converted to negated positives.
+Example: I think you are prepared for the test --> I think you are not unprepared for the test
+## What tasks does it intend to benefit?
+This can act as a valid perturbation that retains the semantics.
+This will benefit for robustness of text classification like sentiment analysis, etc.
+
+
+## What are the limitations of this transformation?
+It is limited to only adjectives and adverbs (all types e.g, comparative and superlative)
+
+## Robustness Evaluation
+
+original accuracy: 96.0  
+dataset_name: 'imdb'  
+model_name: 'aychang/roberta-base-imdb'  
+no_of_examples: 250  
+accuracy after perturbation: 91.0  
+
+The accuracy drops by 5% when tested on imdb dataset for roberta
+

transformations/negated_antonym_perturbation/__init__.py

@@ -0,0 +1 @@
+from .transformation import *

transformations/negated_antonym_perturbation/test.json

@@ -0,0 +1,68 @@
+{
+  "type": "negated_antonym_transformation",
+  "test_cases": [
+    {
+      "class": "NegatedAntonym",
+      "inputs": {
+        "sentence": "I think you are successful."
+      },
+      "outputs": [{
+        "sentence": "I think you are not unsuccessful."
+      }]
+    },
+    {
+      "class": "NegatedAntonym",
+      "inputs": {
+        "sentence": "I think you are prepared for the test."
+      },
+      "outputs": [{
+        "sentence": "I think you are not unprepared for the test."
+      }]
+    },
+    {
+      "class": "NegatedAntonym",
+      "inputs": {
+        "sentence": "He ran quickly."
+      },
+      "outputs": [{
+        "sentence": "He ran not slowly ."
+      }]
+    },
+    {
+      "class": "NegatedAntonym",
+      "inputs": {
+        "sentence": "He plays the flute beautifully."
+      },
+      "outputs": [{
+        "sentence": "He plays the flute not unattractively."
+      }]
+    },
+    {
+      "class": "NegatedAntonym",
+      "inputs": {
+        "sentence": "I think you are joyful."
+      },
+      "outputs": [{
+        "sentence": "I think you are not sorrowful."
+      }]
+    },
+    {
+      "class": "NegatedAntonym",
+      "inputs": {
+        "sentence": "During Harry's fourth year of school, Harry is unwillingly entered as a participant in the Triwizard Tournament, a dangerous yet exciting contest where three champions, one from each participating school, must compete with each other in three tasks in order to win the Triwizard Cup."
+      },
+      "outputs": [{
+        "sentence": "During Harry's fourth year of school, Harry is not willingly entered as a participant in the Triwizard Tournament, a not safe yet exciting contest where three champions, one from each participating school, must compete with each not same in three tasks in order to win the Triwizard Cup."
+      }]
+    },
+    {
+      "class": "NegatedAntonym",
+      "inputs": {
+        "sentence": "Umbridge's incessant and persistent efforts to land him in trouble and the defensive lessons, Harry begins to lose sleep as he constantly receives disturbing dreams about a dark corridor in the Ministry of Magic, followed by a burning desire to learn more."
+      },
+      "outputs": [{
+        "sentence": "Umbridge's incessant and not caducous efforts to land him in trouble and the not offensive lessons, Harry begins to lose sleep as he constantly receives disturbing dreams about a not light corridor in the Ministry of Magic, followed by a burning desire to learn not less."
+      }]
+    }
+  ]
+}

transformations/negated_antonym_perturbation/transformation.py

@@ -0,0 +1,49 @@
+import nltk
+from nltk.corpus import wordnet
+import nltk.tokenize as nt
+
+from interfaces.SentenceOperation import SentenceOperation
+from tasks.TaskTypes import TaskType
+
+''' Class that converts adjectives and adverbs to its negated antonym'''
+
+
+class NegatedAntonym(SentenceOperation):
+    tasks = [TaskType.TEXT_CLASSIFICATION, TaskType.TEXT_TO_TEXT_GENERATION, TaskType.SENTIMENT_ANALYSIS]
+    languages = ["en"]
+    keywords = ["sentiment-analysis", "tokenizer-required", "highly-meaning-preserving"]
+
+
+    def __init__(self, seed=0):
+        super().__init__(seed)
+
+    def generate(self, sentence):
+        return [self.Neg_Antonym(sentence)]
+
+    def Neg_Antonym(self, sentence):
+
+        tokenized_sent = nt.word_tokenize(sentence)
+        pos_sentences = nltk.pos_tag(tokenized_sent)
+
+        for i in range(len(pos_sentences)):
+            antonyms = []
+            if pos_sentences[i][1] == 'JJ' or pos_sentences[i][1] == 'JJR' or pos_sentences[i][1] == 'JJS' or \
+                    pos_sentences[i][1] == 'RB' or pos_sentences[i][1] == 'RBR' or pos_sentences[i][1] == 'RBS':
+                for syn in wordnet.synsets(tokenized_sent[i]):
+
+                    for lm in syn.lemmas():
+
+                        if lm.antonyms():
+                            antonyms.append(lm.antonyms()[0].name())
+
+                if len(antonyms) != 0:
+                    tokenized_sent[i] = 'not ' + antonyms[0]
+
+        for item in range(len(tokenized_sent)):
+            if tokenized_sent[item] == '.' or tokenized_sent[item] == ',' or tokenized_sent[item][0] == "'":
+                tokenized_sent[item - 1] = tokenized_sent[item - 1] + tokenized_sent[item]
+        tokenized_sent = [x for x in tokenized_sent if x != '.']
+        tokenized_sent = [x for x in tokenized_sent if x != ',']
+        tokenized_sent = [x for x in tokenized_sent if x[0] != "'"]
+
+        return ' '.join(tokenized_sent)