Attribute Extraction from Synthetic Web Search Queries

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 401–409, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Attribute Extraction from Synthetic Web Search Queries Marius Pas¸ca Google Inc. Mountain View, California 94043 mars@google.com Abstract The accuracy and coverage of existing methods for extracting attributes of in- stances from text in general, and Web search queries in particular, are limited by two main factors: availability of in- put textual data to which the methods can be applied, and inherent limitations of the underlying assumptions and algo- rithms being used. This paper proposes a weakly-supervised approach for the ac- quisition of attributes of instances from in- put data available in the form of synthetic queries automatically generated from sub- mitted queries. The generated queries al- low for the acquisition of additional at- tributes, leading to extracted lists of at- tributes of higher quality than with com- parable previous methods. 1 Introduction Motivation: The availability of larger textual data sources has allowed research in information ex- traction to shift its focus towards robust methods that require little or no annotated data, operate at large scale with lower computational costs, and acquire open-domain information (Banko and Et- zioni, 2008). The information is usually targeted at three levels of granularity: classes (e.g., giant planets), class elements or instances (e.g., jupiter, uranus, saturn), and relations among instances. Since these types of information would form the backbone of knowledge bases acquired automati- cally from text (Mooney and Bunescu, 2005), their acquisition has received increased attention over recent years. Among other types of relations targeted by vari- ous extraction methods, attributes (e.g., escape ve- locity, diameter and surface gravity) have emerged as one of the more popular types, as they capture quantifiable properties of their respective classes (giant planets) and instances (jupiter). A variety of attribute extraction methods mine textual data sources ranging from unstructured (Tokunaga et al., 2005) or structured (Cafarella et al., 2008) text within Web documents, to human-compiled ency- clopedia (Wu et al., 2008; Cui et al., 2009) and Web search query logs (Alfonseca et al., 2010), attempting to extract, for a given class or instance, a ranked list of attributes that is as comprehen- sive and accurate as possible. The accuracy and coverage of existing methods (Raju et al., 2008; Alfonseca et al., 2010) for extracting attributes of instances are limited by two main factors: avail- ability of input textual data to which the methods can be applied; and inherent limitations of the un- derlying assumptions and algorithms being used. For example, a simple but effective method was proposed in (Pas¸ca and Van Durme, 2007) for ex- tracting attributes of an instance, by applying a small set of extraction patterns (e.g., A of I) to Web search queries (e.g., “escape velocity of jupiter”). If the input set of queries increased, additional candidate attributes would be extracted. Contributions: This paper introduces a weakly- supervised approach for the acquisition of at- tributes of instances from query logs, by auto- matically …

Content truncated to 500 words