Lightly-Supervised Attribute Extraction

Web search engines can greatly benefit from knowledge about attributes of entities present in search queries. In this paper, we introduce lightly-supervised methods for extracting entity attributes from natural language text. Using these methods, we are able to extract large numbers of attributes of different entities at fairly high precision from a large natural language corpus. We compare our methods against a previously proposed pattern-based relation extractor, showing that the new methods give considerable improvements over that baseline. We also demonstrate that query expansion using extracted attributes improves retrieval performance on underspecified information-seeking queries. 1 Attributes in Web Search Web search engines receive numerous queries requesting information, often focused on a specific entity, such as a person, place or organization. These queries are sometimes general requests, such as “bio of George Bush,” or specific requests, such as “new york mayor.” Accurately identifying the entity (new york) or related attributes (mayor) can improve search results in several ways [1]. For example, knowledge of attributes and entities can identify a query as being a factual request [1, 2]. Query expansion using known attributes of the entity can also improve results [3]. Additionally, an engine could suggest alternative queries based on attributes. If a user searches for just “Craig Ferguson” and “shows” is a known attribute of the entity ”Craig Ferguson”, then an alternative query suggestion could be “Craig Ferguson shows” which may guide the user to more informative results. The widely explored technique of pseudo relevance feedback can also benefit from a known list of entities and attributes [4]. Some view entity and attribute extraction as a primary building block for the automatic creation of large scale knowledge bases aimed at addressing these issues [1]. The first step towards improving search results with attributes is to create lists of entities and attributes. Towards that end, we propose new algorithms that, beginning with a small seed set of entities and attributes, learn to extract new entities and attributes from a large corpus of text. We adopt a bootstrapping approach, where the inputs for our learning algorithms are a large unlabeled corpus and the small seed set containing an entity type of interest, such as seed pairs automatically extracted from query logs [1]. The seed pairs are matched against the corpus to create training instances for the learning algorithms. The algorithms exploit a wide range of instance features to alleviate the effects of noise and sparseness. The algorithms produce a large list of entities and associated attributes, which can be directly applied towards improving web search. This paper proceeds as follows. We begin with some background on attribute extraction and web search applications. We then outline our extraction algorithms. Some examples and evaluations of extracted attributes and entities follow.