Research On Transductive Transfer Learning And The Applications

Only on a premise that sufficient training data are available and test data havethe same data distribution with training data, traditional machine learning methodscan yield satisfied results. While the strongly constrainted premise is difficult to befulfiled. Transfer learning methods can impair the constrained limitation effectivelyby applying the transfer knowledge obtained from a source domain to assist learningthe target domain model.As one kind of transfer learning, transductive transfer learning method is usedto achieve transfer tasks when target domain is different from source domain. Thisthesis proposed an EM-based transductive transfer learning method to deal withproblems of lacking labeled target corpus. The EM-based transductive transferlearning method aimed at achieving target task by obtaining transfer knowledgefrom labeled source data and making a combination with unlabeled target data withthe assistance of EM algorithm. The main research work and innovations arepresented as follows:(1) explored a method to build an EM-based transductive transfer learningmodel, which effectively explored the estimating ability of EM algorithm on loglikelihood of hidden variables and learnt auxiliary knowledge from labeled sourcedata to achieve target tasks on the circumstances that the target data are unlabeled.Meanwhile, transfer knowledge content and the methods to obtain transferknowledge and application methods are introduced.(2) studied and implemented an EM-based text classification transfer learningmethod (EMTC) to handle the problem that cannot be solved by traditional textclassification algorithms, which is an document level based version of EM-basedtransductive transfer learning method. The EMTC method learnt a naive bayesianclassifier from source data, taking the classifier parameter as transfer knowledge toassist the text classification task on target domain. The paper built a Chinese textclassification transfer corpus on the basis of traditional Chinese text classificationcorpus. Related experiment results show that the proposed EMTC method waseffective to solve text classification task between different domains.(3) made an research on how to apply transfer learning and classification todeal with Chinese terminology extraction task. Classification based terminologyextraction method utilizes classifier estimating the termhood of extracted candidateterminologies to yield terminologies in documents. This paper proposed anEM-based terminology extraction transfer method (EMTE) which was a phrase levelbased transductive transfer learning method. The experiment result proved our EMTE method can yield an satisfied terminology list even when we lack of labeledtarget data.