Multi-level Alignment for Attribute Extraction in IEPAD

The problem of information extraction (IE) regards automatic generation of extraction programs (also called wrappers). Similar to compiler generator, the core problem is to generate extraction rules. In this paper, we introduce IEPAD (an acronym for Information Extraction based on PAttern Discovery), a system that generalizes extraction patterns from Web pages without user-labeled examples. The system includes a rule generator, which applies sequence mining techniques to discover possible patterns, a rule viewer, which provides an interface for users to see what each pattern can extract, and an extractor, which extracts information from Web pages based on designated extraction rules. To allow ner extraction, multi-level analysis as well as the alignment of multiple records are adopted for attribute extraction. This new track to IE involves no human e ort (to label examples) and content-dependent heuristics. Experiments show that it can achieve 96% retrieval rate with only one training example over 14 popular Web search sites. Among them, ten are able to achieve 100% extraction with less than ve training pages.