MVIFSA: Enhancing Relation Detection in Knowledge Base Question Answering through Multi-View Information Fusion and Self-Attention

The framework of our MVIFSA:

MVIFSA comprises five network layers: a multi-view embedding layer, an information fusion layer, a complex information representation layer, a residual learning layer, and a self-attention layer. The model diagram is as follows.

How to run our code?

Preliminary

You can download the word encoding files required for the experiment from GloVe.

The environment required for the code is in requirements.

Data preprocessing

1. Configure the dataset in config.ini and the maximum length.

2. You can run preprocess.py to preprocess the dataset to get various training and test sets for model training.

Train model.

Model train

The MVIFSA.py file contains the model building and training code. After data preprocessing, you can directly run the file to get the model training results.

Model evaluate

Once you have saved your training model, run the MVIFSA_eval.py file, which allows you to evaluate the trained model.

Thank you for your interest in our work, and feel free to contact the author with any questions you may have.