Fine-grained Web Content Classification via Entity-level Analytics: The Case of Semantic Fingerprinting

Approaching three decades of Web contents being created, the amount of heterogeneous data of diverse provenance becomes seemingly over-whelming and its organization is a "continuous battle" against time. In parallel, business, sociological, political, and media analysts require a structured access to these contents in order to conduct their studies. To this end, concise and - at the same time - efficient engineering methods are required to classify Web contents accordingly. However, the whole task is not as simple as classifying something as A or B, but to assign the most suitable (sub-)category for each Web content based on a fine-grained classification scheme. In practice, the underlying type hierarchies are commonly excerpts of large scale ontologies containing several hundreds or even thousands of (sub-) types decomposed into a few top-level types. Having such a fine-grained type hierarchy, the engineering task of Web content classification becomes out-most challenging. Our main objective in this work is to investigate whether entity-level analytics can be utilized to characterize a Web content and align it onto a fine-grained hierarchy. We hypothesize that "You know a document by the named entities it contains". To this end, we present a novel concept, called "Semantic Fingerprinting" that allows Web content classification solely based on the information derived from the named entities contained in a Web document. It encodes the semantic nature of a Web content into a concise vector, namely the semantic fingerprint. Thus, we expect that semantic fingerprints, when utilized in combination with machine learning, will enable a fine-grained classification of Web contents. In order to empirically validate the effectiveness of semantic fingerprinting, we perform a case study on the classification of Wikipedia documents. Even further, we thoroughly examine the results obtained by analyzing the performance of Semantic Fingerprinting with respect to the characteristics of the data set used for the experiments. In addition, we also investigate performance aspects of the engineered approach by discussing the run-time in comparison with its competitor baselines. We observe that the semantic fingerprinting approach outperforms the state-of-the-art baselines as it raises Web contents to the entity-level and captures their core essence. Moreover, our approach achieves a superior run time performance on the test data in comparison to competitors.