Statistical Learning from Multiple Information Sources

In intelligent information processing tasks such as pattern recognition and information retrieval (IR), probabilistic models are now widely used because they can represent ambiguities of observed data and are robust against noises. Parameters of probabilistic models are statistically estimated (learned) from given training data. However, when the training data contain an insufficient amount of information, the learned model becomes unreliable and its performance severely deteriorates. This thesis proposes two novel learning algorithms that use multiple information sources to mitigate this information scarcity problem in the following two applications. The first application is solving classification problems in which optimal class labels are automatically assigned to observations whose class labels are unknown. Among various types of classification problems, this paper considers classification of sequences that consist of sequential observation points. As a classifier, we focus on the hidden Markov model (HMM), which has been widely used for the classification of sequences. Generally, an HMM is trained on labeled data that consist of observed feature values and class labels. However, due to the high labeling cost, the amount of labeled training data is often small. In this thesis, we propose a learning scheme called semi-supervised learning to improve the classification performance even if the amount of labeled training data is small. The proposed scheme uses both of a small amount of labeled data, and unlabeled data that are not usually used for learning. First, we design a suitable HMM structure for using the unlabeled sequences. Then, we formally derive a semisupervised learning algorithm in which the convergence property is theoretically guaranteed. Next, we apply the proposed method to two types of time series sequences: those acquired from sign language sign data and those acquired from speech phoneme data. Experimental results show that the proposed method outperforms conventional methods. The second application is solving IR problems, especially cross-media IR in which queries and corresponding target data belong to different media. More specifically, this thesis focuses on situations where queries are represented as text and target data are images. When textual annotations explaining the contents of images are provided, such cross-media image retrieval can be regarded as ordinary textual IR. In text retrieval, associations between words can be learned from data and used to relate queries and text in the documents. However, in image retrieval, the number of annotations is too small to learn the word relationships. Using the fact that annotations and images are related, we regard images as the paired data of annotations. ∗Doctor’s Thesis, Department of Information Systems, Graduate School of Information Science, Nara Institute of Science and Technology, NAIST-IS-DT0161006, March 24, 2004.