Using Wikipedia to Bootstrap Open Information Extraction

摘要

By converting unstructured, natural-language text to relational form, information extraction enables many powerful Data Management techniques. However, in order to scale IE to the Web, we must focus on open IE - a paradigm that tackles an unbounded number of relations, eschews domain-specific training data, and scales computationally [2, 11]. This paper describes Kylin, which uses self-supervised learning to train relationally-targeted extractors from Wikipedia infoboxes. We explained how shrinkage and retraining allow Kylin to improve extractor robustness, and we demonstrate that these extractors can successfully mine tuples from a broader set of Web pages. Finally, we argued that the best way to utilize human efforts is by inviting humans to quickly validate the correctness of machine-generated extractions.