Triple based Background Knowledge Ranking for Document Enrichment

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 917–927, Dublin, Ireland, August 23-29 2014. Triple based Background Knowledge Ranking for Document Enrichment Muyu Zhang, Bing Qin∗, Ting Liu, Mao Zheng Research Center for Social Computing and Information Retrieval Harbin Institute of Technology, China {myzhang,qinb,tliu,mzheng}@ir.hit.edu.cn Abstract Document enrichment is the task of retrieving additional knowledge from external resource over what is available through source document. This task is essential because of the phenomenon that text is generally replete with gaps and ellipses since authors assume a certain amount of background knowledge. The recovery of these gaps is intuitively useful for better understanding of document. Conventional document enrichment techniques usually rely on Wikipedia which has great coverage but less accuracy, or Ontology which has great accuracy but less cover- age. In this study, we propose a document enrichment framework which automatically extracts “argument1, predicate, argument2” triple from any text corpus as background knowledge, so that to ensure the compatibility with any resource (e.g. news text, ontology, and on-line ency- clopedia) and improve the enriching accuracy. We first incorporate source document and back- ground knowledge together into a triple based document-level graph and then propose a global iterative ranking model to propagate relevance score and select the most relevant knowledge triple. We evaluate our model as a ranking problem and compute the MAP and P&N score to validate the ranking result. Our final result, a MAP score of 0.676 and P&20 score of 0.417 outperform a strong baseline based on search engine by 0.182 in MAP and 0.04 in P&20. 1 Introduction Document enrichment is the task to acquire background knowledge from external resources and recover the omitted information automatically for certain document. This task is essential because authors usu- ally omit basic but well-known information to make the document more concise. For example, author omits “Baghdad is the captain of Iraqi” in the text of Figure 1 (a), which is well-known to readers. Dur- ing reading process, these gaps will be automatically plugged effortlessly by the background knowledge in human brain. However, the situation is different for machine because it lacks the ability to acquire and select the proper background knowledge, which limits the performances of certain NLP applica- tions. Document enrichment has been proved helpful in these tasks such as web search (Pantel and Fuxman, 2011), coreference resolution (Bryl et al., 2010), document cluster (Hu et al., 2009) and entity disambiguation (Bunescu and Pasca, 2006; Sen, 2012). In the past, there are mainly two kinds of document enrichment researches according to the resource they relying on. The first line of works make use of WikiPedia, the largest available on-line encyclopedia as resource and link the entity (e.g. Baghdad) of document to its corresponding Wiki page (e.g. Baghdad 1 in WikiPedia), so that to enrich the document with the context of Wiki page (Bunescu and Pasca, 2006; Cucerzan, 2007; Han et al., 2011; Kataria et al., 2011; Sen, 2012; He et al., 2013). Despite the great …

Content truncated to 500 words