Skip to content

Latest commit

 

History

History

Abusive text classification

README.md

Abusive Text Classification

Problem Statement

This assignment is divided into 3 parts on the basis pf the machine learning model used for classification of abusive text. The first part is the implementation of the model using the KNN algorithm. The second part is the implementation of the model using the LSTM Neural Network. The third part is the implementation of the model using mBert and MURiL.

Dataset

The dataset contains text messages in Hindi language with labels. The labels are 0 and 1. 0 represents non-abusive text and 1 represents abusive text.

Size of dataset: (20184, 2)

Data Preprocessing

  • Created a set of stop words in Hindi languages from several resources.
  • Removed the stop words from the dataset sentences
  • Removed punctuation marks from sentences
  • Converted emojis to text equivalent representation using emot library
  • Removed digits from text

Part 1: KNN

  • Used the preprocessed data to tokenize and calculate tf-idf values using TfIdfVectorizer
  • Split dataset into 80:20 split for training and testing
  • Fitted the train data and tested on test data

Part 2: LSTM

  • Created a class for LSTM architecture with the following layers, activation function, and dimensions 
    LSTM(
    (embedding): Embedding(29941, 300)
    (lstm): LSTM(300, 600, num_layers=2, batch_first=True, dropout=0.3)
    (fc): Linear(in_features=600, out_features=1, bias=True)
    (dropout): Dropout(p=0.3, inplace=False)
    (sig): Sigmoid()
)
  • Created vectorized dataset using word_tokenizer()
  • Padded all the sentences to a maximum length
  • Split the dataset into an 80:20 ratio
  • Trained the model with the below hyper-parameters 
    vocab_size = len(vocab)
embedding_dim = 300
hidden_dim = 600
num_layers = 2
epochs = 10
lr = 0.001
  • Also embedded the logic of early stopping by maintaining a counter.

Part 3: mBert and MURiL

  • Tokenized and encoded the dataset using hugging face's “bert-base-multilingual-cased” and “google/muril-base-cased” tokenizer.
  • Fine-tuned prebuilt model for the same mBert and MURiL architecture

Results

Model Accuracy Macro F1 Score
KNN (k=18) 64 62
LSTM 78.30 78.09
mBert 82.44 82.17
MURiL 85.29 84.98