Given a question, related paragraphs from a Wikipedia article (in the shuffle order), the task is finding paragraph which answers the question of each test case.

F1 measure based on precision and recall is used to rank competing submissions.

For detailed, visit: https://challenge.zalo.ai/portal/question-answering

• train data: 18108 question text pairs
• test data: 2678 question text pairs

Some interesting/confusing/mistake question text pairs:

We used handcrafted features, bert and combined on original data (only use question, text), or add title data, add transitivity rules and add external data (Squad 2.0 translated to Vietnamese).

We have made 73 single models from the above combinations.

We selected the best model to submit. Although ensemble version have the better result, however the predict time is so much longer than one model, so one model is submit. The config is below:

Init checkpoint at https://github.com/lampts/bert4vn (BERT for Vietnamese)

num_train_epochs = 5.0

max_seq_length = 512 

train_batch_size = 128

learning_rate = 2e-5

Please download the init checkpoint at here and put into the folder checkpoint.

Put data in the folder /data/

Run command sh train.sh

If you want to only predict, please download the checkpoint of trained model at here

Run command: sh predict.sh.

ALBERT for Vietnamese will be release at https://github.com/ngoanpv/albert_vi