Cyberbullying Detection

Description

Identify cyberbullying using a multi-class classification framework that distinguishes six different types of cyberbullying. We have used a Twitter dataset from Kaggle and applied various techniques such as text cleaning, data augmentation, document assembling, universal sentence encoding and tensorflow classification model to process and analyze the data. We have also used snscrape to retrieve tweet data for validating our model’s performance. Our results show that our model achieved an accuracy of 85% for testing data and 89% for training data.

Live Demo Open in Colab Download Copy S3 URI

How to use

documentAssembler = DocumentAssembler()\
    .setInputCol("cleaned_text")\
    .setOutputCol("document")

use = UniversalSentenceEncoder.pretrained(name="tfhub_use_lg", lang="en")\
 .setInputCols("document")\
 .setOutputCol("sentence_embeddings")\
 .setDimension(768)

classifierdl = ClassifierDLApproach()\
  .setInputCols(["sentence_embeddings"])\
  .setOutputCol("class")\
  .setLabelColumn("cyberbullying_type")\
  .setBatchSize(16)\
  .setMaxEpochs(42)\
  .setDropout(0.4) \
  .setEnableOutputLogs(True)\
  .setLr(4e-3)
use_clf_pipeline = Pipeline(
    stages = [documentAssembler,
        use,
        classifierdl])

Results

           precision    recall  f1-score   support

                age       0.94      0.96      0.95       796
          ethnicity       0.94      0.94      0.94       810
             gender       0.87      0.86      0.86       816
  not_cyberbullying       0.74      0.67      0.70       766
other_cyberbullying       0.67      0.71      0.69       775
           religion       0.94      0.96      0.95       731

           accuracy                           0.85      4694
          macro avg       0.85      0.85      0.85      4694
       weighted avg       0.85      0.85      0.85      4694

Model Information

Included Models

• DocumentAssembler
• UniversalSentenceEncoder
• ClassifierDLModel