Synthesis of forgiving data extractors

We address the problem of synthesizing a robust data-extractor from a family of websites that contain the same kind of information. This problem is common when trying to aggregate information from many web sites, for example, when extracting information for a price-comparison site. Given a set of example annotated web pages from multiple sites in a family, our goal is to synthesize a robust data extractor that performs well on all sites in the family (not only on the provided example pages). The main challenge is the need to trade off precision for generality and robustness. Our key contribution is the introduction of forgiving extractors that dynamically adjust their precision to handle structural changes, without sacrificing precision on the training set. Our approach uses decision tree learning to create a generalized extractor and converts it into a forgiving extractor, in the form of an XPath query. The forgiving extractor captures a series of pruned decision trees with monotonically decreasing precision, and monotonically increasing recall, and dynamically adjusts precision to guarantee sufficient recall. We have implemented our approach in a tool called TREEX and applied it to synthesize extractors for real-world large scale web sites. We evaluate the robustness and generality of the forgiving extractors by evaluating their precision and recall on: (i) different pages from sites in the training set (ii) pages from different versions of sites in the training set (iii) pages from different (unseen) sites. We compare the results of our synthesized extractor to those of classifier-based extractors, and pattern-based extractors, and show that TREEX significantly improves extraction accuracy.