Semi-Supervised Structured Output Learning

Опубликовано 12 августа 2016, 2:10

Partial label scenarios arise commonly in web applicationssuch as hierarchical classification, multi-label classification and informationextraction from web pages. For example, label information may be available onlyat the internal node level (not at the leaf level) for some pages in ahierarchical classification problem. In a multi-label classification problem,it may be available only for some of the classes (in each example). Similarly,in a sequence learning problem, we may have label information only for somenodes in the training sequences. Conventionally, marginal likelihood maximization technique has been used to solve these problems. In such asolution unlabeled examples and domain knowledge are not used. Domain knowledgeplays a key role in achieving improved performance with unlabeled data. In thistalk we discuss probabilistic structured output models, and the mechanisms bywhich partially/unlabeled examples can be used with domain knowledge forlearning. Experiments on real-life hierarchical and multi-label learningproblems show that significant improvements in accuracy are achieved byincorporating regularizations based on domain knowledge and other forms (e.g,entropy), when most of the examples are only partially labeled.