This repository contains our code for the Microsoft AI Challenge India 2018. Our solution is Ranked 5th on Public Leaderboard giving 0.7114 MRR Score.
Microsoft organized Microsoft AI Challenge 2018 for India in November 2018. The problem statement was to rank the right response to queries(web) given 10 possible responses.
- Fine-tuning BERT with different sequence length
- Ensembling results in different ways
Total Epochs | Model Decription (seq. length) |
MRR |
---|---|---|
1 | 120 | 0.6909 |
1 | 160 | 0.7035 |
2 | 120 + 120 | 0.6999 |
2 | 120 + 160 | 0.7073 |
3 | 120 + 120 + 120 | 0.6981 |
3 | 120 + 160 + 198 | 0.7107 |
4 | 120 + 120 + 120 + 120 | 0.6899 |
Fine-tune an instance of BertForMultipleChoice by Hugging Face