Combination of Neural Networks for Multi-label Document Classification
in 22th International Conference on Applications of Natural Language to Information Systems (NLDB 2017) (2017)
This paper deals with multi-label classification of Czech documents using several combinations of neural networks.
It is motivated by the assumption that different nets can keep some complementary information and that it should be useful to combine them.
The main contribution of this paper consists in a~comparison of several combination approaches to improve the results of the individual neural nets.
The experimental results show that it is useful to combine the individual nets to improve the final score of multi-label classification.
We further show that the results of all the combination approaches outperform the individual nets and are comparable.
However, the best combination method is the supervised one which uses a~feed-forward neural net with sigmoid activation function.
We also analyze the results in order to further improve the final accuracy of the whole approach.