Skip to content

Bi-Directional Attention Flow (BiDAF) question answering model enhanced by multi-layer convolutional neural network character embeddings.

License

Notifications You must be signed in to change notification settings

tomassykora/bidaf-question-answering

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

16 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

bidaf-question-answering

A Bi-Directional Attention Flow (BiDAF) question answering model enhanced by 2 convolutional network layers with batch normalisation and a residual connection for the character embeddings as the input to the bidirectional LSTM layer.

Proposed improvements of 4.94 and 5.11 percentage points in EM and F1 scores respectively are described in the paper (although there was only one experiment with the batch norm. layer which means many improvements ahead!).

Character embedding layers EM F1
course baseline 58.23 61.44
1-layer cnn emb + residual 59.57 62.90
2-layer cnn emb + residual 60.46 63.65
2-layer cnn emb + residual + batch norm. 63.17 66.55

This project is originally from the CS224n Stanford course (winter 2020), but solved for the course ZPJa - Natural Language Processing at Brno University of Technology, Faculty of Information Technologies.

Based on the paper - Bidirectional Attention Flow for Machine Comprehension (2018).

About

Bi-Directional Attention Flow (BiDAF) question answering model enhanced by multi-layer convolutional neural network character embeddings.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages