Traditional supervised machine learning requires labeled data for a specific problem of interest. There have been many attempts to reduce this requirement such as approaches based on semi-supervised learning. In recent years, people have started to consider a new strategy known as transfer learning, where labeled data from an old problem (called the source task) is used to assist the learning of a new but related problem (the target task).;In this thesis, we mainly consider an extreme case of transfer learning that we denote as heterogeneous transfer learning - where the feature spaces of the source task and the target tasks are disjoint. We first consider the cross-lingual text classification task, where we need to train a classifier for Chinese but we only have labeled data in English. We adapt the structural correspondence learning (SCL) algorithm for the problem. Furthermore, we generalize the SCL algorithm as a multi-task transfer learning strategy and propose the use of a restricted Boltzmann machine (RBM), a special type of probabilistic graphical models, as an implementation. We also give some preliminary theoretical analysis for the strategy by combining previous work on general transfer learning and multi-task learning.;Finally, we study the problem of co-reference resolution using another kind of graphical models, the conditional random field (CRF). We show that a previously proposed ranking approach, which produces state of the art results, can be viewed as a special case of the model. We go on to show how using a CRF allows us to easily incorporate other NLP tasks such as non-anaphoric identification and noun phrase boundary detection.
展开▼
