/BERT-Dialogue_Classification

A dialogue classifier from the show The Office. This code utilises the Bert (Base) model to classify the dialogues as Jim or Dwight's Dialogues.

Primary LanguageJupyter Notebook

Dialogue Classification using Bidirectional Encoder Representations from Transformers (BERT)

A dialogue classifier from the show The Office. This code utilises the fine-tuned BERT-based model to classify the dialogues as Jim or Dwight's Dialogues.

Jim vs. Dwight Dialogue Speaker Classification

This project focuses on classifying speakers in TV series dialogues using a fine-tuned BERT-based model. The model is trained on a dataset containing dialogues from the TV series "The Office" and can predict the speaker of the line with a 84% accuracy.

Dataset

The dataset used in this project consists of dialogues from The Office. Note: The model was only trained on the dialogues given in the file train.csv for a Kaggle competition by IEEE NITK's Computer Intelligence Society.

Requirements

  • TensorFlow
  • Transformers
  • Pandas
  • NumPy

Application

  1. About the dataset:

    • The training dataset was in CSV format with columns: "id", "line", "speaker".
    • The validation dataset was split from the train.csv file, hence it is also in the the same format.
    • Data was properly cleaned and pre-processed.

  2. Fine tuning the BERT model:

    • Run the BERT-ieeekagglecup-2023.ipynb notebook to fine-tune the BERT model on the training dataset.
    • The notebook tokenizes the text, prepares input tensors, and trains the model.

  3. Evaluation:

    • The trained model is evaluated on the validation dataset using accuracy and classification report metrics.
    • The evaluation results can be used to assess the model's performance.

  4. Making predictions:

    • Use the trained model to predict the speaker of dialogue lines.
    • test.csv dataset is in CSV format with columns: "id", "line".

  5. Generated the submission CSV for the Kaggle Contest - my first one :)

Feel free to contribute, open issues, or submit pull requests to enhance the project!