Instance-Driven Attachment of Semantic Annotations over Conceptual Hierarchies

Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics, pages 503–513, Avignon, France, April 23 - 27 2012. c⃝2012 Association for Computational Linguistics Instance-Driven Attachment of Semantic Annotations over Conceptual Hierarchies Janara Christensen∗ University of Washington Seattle, Washington 98195 janara@cs.washington.edu Marius Pas¸ca Google Inc. Mountain View, California 94043 mars@google.com Abstract Whether automatically extracted or human generated, open-domain factual knowledge is often available in the form of semantic annotations (e.g., composed-by) that take one or more specific instances (e.g., rhap- sody in blue, george gershwin) as their ar- guments. This paper introduces a method for converting flat sets of instance-level annotations into hierarchically organized, concept-level annotations, which capture not only the broad semantics of the desired arguments (e.g., ‘People’ rather than ‘Loca- tions’), but also the correct level of general- ity (e.g., ‘Composers’ rather than ‘People’, or ‘Jazz Composers’). The method refrains from encoding features specific to a partic- ular domain or annotation, to ensure imme- diate applicability to new, previously un- seen annotations. Over a gold standard of semantic annotations and concepts that best capture their arguments, the method sub- stantially outperforms three baselines, on average, computing concepts that are less than one step in the hierarchy away from the corresponding gold standard concepts. 1 Introduction Background: Knowledge about the world can be thought of as semantic assertions or anno- tations, at two levels of granularity: instance level (e.g., rhapsody in blue, tristan und isolde, george gershwin, richard wagner) and concept level (e.g., ‘Musical Compositions’, ‘Works of Art’, ‘Composers’). Instance-level annotations correspond to factual knowledge that can be found in repositories extracted automatically from text (Banko et al., 2007; Wu and Weld, 2010) ∗Contributions made during an internship at Google. or manually created within encyclopedic re- sources (Remy, 2002). Such facts could state, for instance, that rhapsody in blue was composed- by george gershwin, or that tristan und isolde was composed-by richard wagner. In compar- ison, concept-level annotations more concisely and effectively capture the underlying semantics of the annotations by identifying the concepts cor- responding to the arguments, e.g., ‘Musical Com- positions’ are composed-by ‘Composers’. The frequent occurrence of instances, relative to more abstract concepts, in Web documents and popular Web search queries (Barr et al., 2008; Li, 2010), is both an asset and a liability from the point of view of knowledge acquisition. On one hand, it makes instance-level annotations rel- atively easy to find, either from manually created resources (Remy, 2002; Bollacker et al., 2008), or extracted automatically from text (Banko et al., 2007). On the other hand, it makes concept- level annotations more difficult to acquire di- rectly. While “Rhapsody in Blue was composed by George Gershwin [..]” may occur in some form within Web documents, the more abstract “Musical compositions are composed by musi- cians [..]” is unlikely to occur. A more practical approach to collecting concept-level annotations is to indirectly derive them from already plenti- ful instance-level annotations, effectively distill- ing factual knowledge into more abstract, concise and generalizable …

Content truncated to 500 words