Welcome to our submission for the COMP34812 shared task on Natural Language Inference (NLI). Our project implements and compares two distinct deep learning approaches for pairwise sequence classification. Project Overview
Approach B: BiLSTM
- A bidirectional LSTM network that captures contextual dependencies in both directions.
- Model Card: View Model Card
- Demo Notebook: NLU_Method_B/BILSTM_Demo.ipynb
- Training Notebook: NLU_Method_B/BILSTM_RNN_TrainandEvaluate.ipynb
Approach C: Transformer Ensemble
- An ensemble of transformer-based models leveraging the strength of multiple pre-trained language models.
- Model Card: View Model Card
- Demo Notebook: NLU_Method_C/NLI_Transformer_demo.ipynb
- Training Notebook: NLU_Method_C/Transformer_Train_and_Evaluate.ipynb
Our approach and results are presented on our poster: NLU Presentation Poster.pdf
Both demo scripts are designed for Google Colab and will automatically install required Python libraries.
Steps to run:
- Open the relevant notebook in Colab.
- Follow the instructions at the top to upload your test dataset.
- Run the notebook cells to generate predictions.
BiLSTM Demo: NLU_Method_B/BILSTM_Demo.ipynb
Transformer Demo: NLU_Method_C/NLI_Transformer_demo.ipynb
Outputs will be in the required CSV format for submission.
Our training notebooks are also Colab-friendly and fully documented.
BiLSTM Training: NLU_Method_B/BILSTM_RNN_Trainer.ipynb
Transformer Training: NLU_Method_C/Transformer_Train_and_Evaluate.ipynb
Both scripts handle:
- Dataset loading
- Model training
- Evaluation on the dev set
- Saving models for reuse in demo notebooks
We used the official dataset provided in the coursework:
Our trained models are stored on Hugging Face for easy reuse: