A Multiple Instance Learning Algorithm Using Graph Convolutional Network for Speech Content Classification

Recently, speech content classification begins to be formulated as a multiple instance learning (MIL) problem by some researchers. Under the MIL framework, a speech signal is considered as a bag with labels, and speech fragments are regarded as instances without corresponding labels. On this basis, we propose a new algorithm, named MI-CNN-G, for speech content classification using Graph Convolutional Network (GCN) under the MIL framework. First, we only train the MIL classifier, whose architecture is based on Convolutional Neural Network, for bag classification. Second, we utilize GCN to handle graphs composed of instance features that are the outputs of the MIL classifier for correcting the predicted probabilities of the classifier and then updated the classifier and GCN together. Finally, we employ the MI-CNN-G model to predict the labels of unknown bags. The experimental results show that GCN is effective in improving the performance of the MIL classifier on speech content classification and the MI-CNN-G model outperforms the baseline approaches in the Average Precision metric.