Quality Impact of Value Matching and Scoring in Top-k Entity Attribute Extraction∗

The entity attribute extraction problem, or how to extract entities and their attribute values from natural language Web documents, is of critical importance for Web search and information access in general. Unfortunately, because of the noisy nature of theWeb and its scale, entity attribute extraction is notoriously challenging in terms of both extraction efficiency and quality. In our earlier work [24], we proposed a top-k extraction processing approach that addressed the efficiency challenge: Our approach leveraged a popularitybased scoring function to rank Web pages according to their entity-specific importance, and focused the extraction effort over the highly ranked pages for each entity of interest. The extraction quality resulting from this efficiency-motivated extraction approach, however, has not been studied and is the focus of this paper. Specifically, we make progress toward addressing the quality challenge through an in-depth analysis of two critical components of the extraction process, namely, matching and scoring of extracted attribute values. The design choices for these components can substantially impact the quality of the entity attribute extraction process, as we demonstrate with experiments with a state-of-the-art extraction system and entities from two domains of interest.