Natural Language Inference Model

• 

  Here is a demo of our webapp showing the predicted and expected relationship between pairs of sentences (premise and hypothesis).

Inspiration

Getting Started competitions are semi-permanent competitions on Kaggle that are meant to be used by new users just starting in the field of machine learning. They are designed to be good first or second ML projects. The one we chose is called “Contradictory, My Dear Watson” and is about detecting contradiction and entailment in multilingual text. We selected this competition because it seemed more challenging than the other ones from the same category and because the subject was somewhat related to cognitive science (our program). For more details about the competition: https://www.kaggle.com/c/contradictory-my-dear-watson.

What it does

If you have two sentences, there are three ways they could be related: one could entail the other, one could contradict the other, or they could be unrelated. Natural Language Inferencing (NLI) is a popular NLP problem that involves determining how pairs of sentences (consisting of a premise and a hypothesis) are related. The goal of this project is to create a NLI model that assigns labels of 0, 1, or 2 (corresponding to entailment, neutral, and contradiction) to pairs of premises and hypotheses. To make things more interesting, the train and test set include text in fifteen different languages.

How we built it

To realize this project, we followed the general guidelines of the starter notebook suggested in the Kaggle competition. We used a pre-trained BERT, fined-tuned over the dataset from the competition. To reduce the time required for the training/fine-tuning phase, we made use of Tensor Processing Units (TPUs). Finally, we built a webapp to showcase the results of our NLI model.

Built With

• bert
• javascript
• python
• tensorflow
• tpu