Recent advances in sensors, digital storage, computing and communications technologies have led to a proliferation of autonomously operated, geographically distributed data repositories in virtually every area of human endeavor, including e-business and e-commerce, e-science, e-government, security informatics, etc. Effective use of such data in practice (e.g., building useful predictive models of consumer behavior, discovery of factors that contribute to large climatic changes, analysis of demographic factors that contribute to global poverty, analysis of social networks, or even finding out what makes a book a bestseller) requires accessing and analyzing data from multiple heterogeneous sources. The Semantic Web enterprise (Berners-Lee et al., 2001) is aimed at making the contents of the Web machine interpretable, so that heterogeneous data sources can be used together. Thus, data and resources on the Web are annotated and linked by associating meta data that make explicit the ontological commitments of the data source providers or, in some cases, the shared ontological commitments of a small community of users. Given the autonomous nature of the data sources on the Web and the diverse purposes for which the data are gathered, in the absence of a universal ontology it is inevitable that there is no unique global interpretation of the data, that serves the needs of all users under all scenarios. Many groups have attempted to develop, with varying degrees of success, tools for flexible integration and querying of data from semantically disparate sources (Levy, 2000; Noy, 2004; Doan, & Halevy, 2005), as well as techniques for discovering semantic correspondences between ontologies to assist in this process (Kalfoglou, & Schorlemmer, 2005; Noy and Stuckenschmidt, 2005). These and related advances in Semantic Web technologies present unprecedented opportunities for exploiting multiple related data sources, each annotated with its own meta data, in discovering useful knowledge in many application domains. While there has been significant work on applying machine learning to ontology construction, information extraction from text, and discovery of mappings between ontologies (Kushmerick, et al., 2005), there has been relatively little work on machine learning approaches to knowledge acquisition from data sources annotated with meta data that expose the structure (schema) and semantics (in reference to a particular ontology). However, there is a large body of literature on distributed learning (see (Kargupta, & Chan, 1999) for a survey). Furthermore, recent work (Zhang et al., 2005; Hotho et al., 2003) has shown that in addition to data, the use of meta data in the form of ontologies (class hierarchies, attribute value hierarchies) can improve the quality (accuracy, interpretability) of the learned predictive models. The purpose of this chapter is to precisely define the problem of knowledge acquisition from semantically heterogeneous data and summarize recent advances that have led to a solution to this problem (Caragea et al., 2005).
An Architecture for Data and Knowledge Acquisition for the Semantic Web: The AGROVOC Use Case