An extraction-rule generation and training system uses information obtained from multiple markup language documents (e.g. web pages) of similar structure to generate an extraction rule for extracting datapoints from markup language documents. By using information extracted from multiple documents of similar structure, including information regarding correlations between such documents, the method produces data extraction rules that provide improved datapoint extraction reliability. Where the structures of two or more documents are not sufficiently similar, the system maintains separate extraction rules for the same datapoint, and applies these separate extraction rules in combination to particular markup language documents to extract the datapoint.