Introduction

• This is the Information Retrieval HW6.
• Using BM25 and transformer-based language model to compute the relation between given querys and documents.

Usage

• All file have argument parser, please use python3 FILENAME -h to get more.

Approach

• Use transformer-based doing multiple choice task.

preprocess

• [CLS] query [SEP] document [SEP]
• Token length truncate to 512 by tokenizer.

Hyperparameter

• Pretrained lenguage model: bert-base-uncased
• Optimizer: AdamW
• Learning rate = 1e-5
• Split 20 documents of training queries to grid search optimal 𝜶 for BERT
• Batch size = 3
• Num. epochs = 2
• Num. of negative documents = 3
• 𝜶 from greedy search = 1.13