Learning with Taxonomies: Classifying Documents and Words

Automatically extracting semantic information about word meaning and document topic from text typically involves an extensive number of classes. Such classes may represent predefined word senses, topics or document categories and are often organized in a taxonomy. The latter encodes important information, which should be exploited in learning classifiers from labeled training data. To that extent, this paper presents an extension of multiclass Support Vector Machine learning which can incorporate prior knowledge about class relationships. The latter can be encoded in the form of class attributes, similarities between classes or even a kernel function defined over the set of classes. The paper also discusses how to specify and optimize meaningful loss functions based on the relative position of classes in the taxonomy. We include experimental results for text categorization and for word sense classification.