Outclassing Wikipedia in Open-Domain Information Extraction: Weakly-Supervised Acquisition of Attributes over Conceptual Hierarchies

Proceedings of the 12th Conference of the European Chapter of the ACL, pages 639–647, Athens, Greece, 30 March – 3 April 2009. c⃝2009 Association for Computational Linguistics Outclassing Wikipedia in Open-Domain Information Extraction: Weakly-Supervised Acquisition of Attributes over Conceptual Hierarchies Marius Pas¸ca Google Inc. Mountain View, California 94043 mars@google.com Abstract A set of labeled classes of instances is ex- tracted from text and linked into an exist- ing conceptual hierarchy. Besides a signif- icant increase in the coverage of the class labels assigned to individual instances, the resulting resource of labeled classes is more effective than similar data derived from the manually-created Wikipedia, in the task of attribute extraction over con- ceptual hierarchies. 1 Introduction Motivation: Sharing basic intuitions and long- term goals with other tasks within the area of Web- based information extraction (Banko and Etzioni, 2008; Davidov and Rappoport, 2008), the task of acquiring class attributes relies on unstructured text available on the Web, as a data source for ex- tracting generally-useful knowledge. In the case of attribute extraction, the knowledge to be ex- tracted consists in quantifiable properties of var- ious classes (e.g., top speed, body style and gas mileage for the class of sports cars). Existing work on large-scale attribute extraction focuses on producing ranked lists of attributes, for target classes of instances available in the form of flat sets of instances (e.g., ferrari modena, porsche carrera gt) sharing the same class label (e.g., sports cars). Independently of how the input target classes are populated with instances (man- ually (Pas¸ca, 2007) or automatically (Pas¸ca and Van Durme, 2008)), and what type of textual data source is used for extracting attributes (Web docu- ments or query logs), the extraction of attributes operates at a lexical rather than semantic level. Indeed, the class labels of the target classes may be not more than text surface strings (e.g., sports cars) or even artificially-created labels (e.g., Car- toonChar in lieu of cartoon characters). More- over, although it is commonly accepted that sports cars are also cars, which in turn are also motor ve- hicles, the presence of sports cars among the input target classes does not lead to any attributes being extracted for cars and motor vehicles, unless the latter two class labels are also present explicitly among the input target classes. Contributions: The contributions of this paper are threefold. First, we investigate the role of classes of instances acquired automatically from unstructured text, in the task of attribute extrac- tion over concepts from existing conceptual hi- erarchies. For this purpose, ranked lists of at- tributes are acquired from query logs for various concepts, after linking a set of more than 4,500 open-domain, automatically-acquired classes con- taining a total of around 250,000 instances into conceptual hierarchies available in WordNet (Fell- baum, 1998). In comparison, previous work extracts attributes for either manually-specified classes of instances (Pas¸ca, 2007), or for classes of instances derived automatically but considered as flat rather than hierarchical classes, and manually associated to existing semantic concepts (Pas¸ca and …

Content truncated to 500 words