Retrieval Term Prediction Using Deep Learning Methods

This paper presents methods to predict retrieval terms from relevant/surrounding words or descriptive texts in Japanese by using deep learning methods, which are implemented with stacked denoising autoencoders (SdA), as well as deep belief networks (DBN). To determine the effectiveness of using DBN and SdA for this task, we compare them with conventional machine learning methods, i.e., multi-layer perceptron (MLP) and support vector machines (SVM). We also compare their performance in case of using three regularization methods, the weight decay (L2 regularization), sparsity (L1 regularization), and dropout regularization. The experimental results show that (1) adding automatically gathered unlabeled data to the labeled data for unsupervised learning is an effective measure for improving the prediction precision, and (2) using DBN or SdA results in higher prediction precision than using SVM or MLP, whether or not regularization methods are used.