Encoding Distributional Semantics into Triple-Based Knowledge Ranking for Document Enrichment

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 524–533, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Encoding Distributional Semantics into Triple-Based Knowledge Ranking for Document Enrichment Muyu Zhang1∗, Bing Qin1, Mao Zheng1, Graeme Hirst2, and Ting Liu1 1Research Center for Social Computing and Information Retrieval Harbin Institute of Technology, Harbin, China 2Department of Computer Science, University of Toronto, Toronto, ON, Canada {myzhang,qinb,mzheng,tliu}@ir.hit.edu.cn gh@cs.toronto.edu Abstract Document enrichment focuses on retriev- ing relevant knowledge from external re- sources, which is essential because text is generally replete with gaps. Since conven- tional work primarily relies on special re- sources, we instead use triples of Subject, Predicate, Object as knowledge and in- corporate distributional semantics to rank them. Our model first extracts these triples automatically from raw text and converts them into real-valued vectors based on the word semantics captured by Latent Dirich- let Allocation. We then represent these triples, together with the source document that is to be enriched, as a graph of triples, and adopt a global iterative algorithm to propagate relevance weight from source document to these triples so as to select the most relevant ones. Evaluated as a rank- ing problem, our model significantly out- performs multiple strong baselines. More- over, we conduct a task-based evaluation by incorporating these triples as additional features into document classification and enhances the performance by 3.02%. 1 Introduction Document enrichment is the task of acquiring rel- evant background knowledge from external re- sources for a given document. This task is essen- tial because, during the writing of text, some ba- sic but well-known information is usually omitted by the author to make the document more concise. For example, Baghdad is the capital of Iraq is omitted in Figure 1a. A human will fill these gaps automatically with the background knowledge in his mind. However, the machine lacks both the ∗ This work was partly done while the first author was visiting University of Toronto. necessary background knowledge and the ability to select. The task of document enrichment is pro- posed to tackle this problem, and has been proved helpful in many NLP tasks such as web search (Pantel and Fuxman, 2011), coreference resolu- tion (Bryl et al., 2010), document cluster (Hu et al., 2009) and entity disambiguation (Sen, 2012). We can classify previous work into two classes according to the resources they rely on. The first line of work uses Wikipedia, the largest on-line en- cyclopedia, as a resource and introduces the con- tent of Wikipedia pages as external knowledge (Cucerzan, 2007; Kataria et al., 2011; He et al., 2013). Most research in this area relies on the text similarity (Zheng et al., 2010; Hoffart et al., 2011) and structure information (Kulkarni et al., 2009; Sen, 2012; He et al., 2013) between the mention and the Wikipedia page. Despite the apparent suc- cess of these methods, most Wikipedia pages con- tain too much information, most of …

Content truncated to 500 words