Automatically Training Form Classifiers

Abstract

The state-of-the-art in domain-specific Web form discovery relies on supervised methods requiring substantial human effort in providing training examples, which limits their applicability in practice. This paper proposes an effective alternative to reduce the human effort: obtaining high-quality domain-specific training forms. In our approach, the only user input is the domain of interest; we use a search engine and a focused crawler to locate query forms which are fed as training data into supervised form classifiers. We tested this approach thoroughly, using thousands of real Web forms from six domains, including a representative subset of a publicly available form base to validate this approach. The results reported in this paper show that it is feasible to mitigate the demanding manual work required by some methods of the current state-of-the-art in form discovery, at the cost of a negligible loss in effectiveness.