Selective Retrieval for Categorization of Semi-structured Web Resources

摘要

A typical on-line content directory contains factual information about entities (e.g., address of a company) together with entity categories (e.g., company's industries). The categories are a salient element of the system as they allow users to browse for entities of a chosen type. Assigning categories manually can be a challenging task, considering that an entity can belong to few out of hundreds of categories (e.g., all possible industry types). Instead we suggest to augment this process with an automatic categorization system that suggests categories based on the entity's home page. To improve the accuracy of results, the system follows links extracted from the home page and uses retrieved content to expand an entity's term profile. The profile is later used by a multi-label classification system to assign categories to the entity. The key element of the system is a link ranking module, which uses home page features (e.g., position and anchor text of links) to select links that are most likely to improve the categorization results. Evaluation on a data set of nearly ten thousand company home pages confirmed that the link ranking approach allows the system to limit the retrieval and processing costs to allow real-time responses and still outperform the categorization results of baseline systems.