Natural Language Inference (NLI)

• This project's scope is to build a ZeroShot Cross-lingual Natural Language Inference (XNLI). NLI is a pair sequences classification task, to classify the logical relationship between 2 sentences, whether it is entailment, contradiction, neutral.
• Data sets provided for this task where loaded using Huggingface nlp using both MNLI Multi-genre NLI English corpus for both training and validation data sets, and using cross-lingual NLI corpus XNLI made up of 15 languages (spanning over different language families).
  • dataset did not suffer any labels imbalance
  • data sets parser is found in hw3/stud/data_loader directory, different parsers were created for different models used
• Different models were used starting from BiLSTM model up to using Pretrained language models
  1. Baseline model
  2. K_model
  3. mBERT
  4. XLM model based on either (XLM) or (XLM & TLM)
  5. XLM-RoBERTa
  • Models can be found in hw3/stud/models/model_architectures.py and hyperparameters script is in the same directory

---

To use this repo

• preferably create virtual env

conda create -n xnli_task python=3.7
conda activate xnli_task

• Install requirements

pip install -r requirements.txt

• To train the models

python3 hw3/stud/main_baseline.py  # for baseline or K model
python3 hw3/stud/main_bert.py  # for mBERT model
python3 hw3/stud/main_xlm.py  # for XLM model
python3 hw3/stud/main_xlm-r.py  # for XLM-R model 

• Check XNLI_XLMR.ipynb notebook for commented code, added to the scripts
• Check report for more theoritical insights report.pdf