preprocessing:
  steps:
    - lowercase
    - remove_urls (http/https, www, pic.twitter, youtube, facebook, hungama, attached)
    - remove_mentions
    - remove_hashtags
    - normalize_elongated_words
    - remove_html_entities
    - normalize_whitespace
    - remove_duplicate_rows
  noise_removed:
    duplicate_texts: 11
    texts_with_urls: 458
    texts_with_elongated_words: 2176
  result:
    rows_before: 29550
    rows_after: 29506
    retained_features: [clean_text, hate_label, language, text_length, word_count]
tokenization:
  bilstm_lstm: keras Tokenizer + pad_sequences
  embeddings: pretrained GloVe / Word2Vec / FastText embedding matrix