N-gram Fragment Sequence Based Unsupervised Domain-Specific Document Readability

Proceedings of COLING 2012: Technical Papers, pages 1309–1326, COLING 2012, Mumbai, December 2012. N-gram Fragment Sequence Based Unsupervised Domain-Specific Document Readability Shoaib Jameel Xiaojun Qian Wai Lam Department of Systems Engineering and Engineering Management, The Chinese University of Hong Kong. ���������� ������� �������������������� ABSTRACT Traditional general readability methods tend to underperform in domain-specific document retrieval because they fail to effectively differentiate the reading difficulty of the individual domain-specific terms and the semantic associations between the textual units in a document. On the other hand, recently proposed domain-specific readability methods have relied upon an external knowledge base which may be unavailable in some domains. We develop a novel un- supervised framework for computing domain-specific document readability. Our model does not require an ontology or a knowledge base to capture domain-specific terms in a document. The sequential flow of terms in a document is modeled as a connected sequence of n-gram fragments in the latent concept space. We investigate an automatic sequential n-gram deter- mination scheme that aids in capturing appropriate n-gram fragments which are semantically associated with the document’s theme and cohesive with the context. The domain-specific readability cost of a document is computed based on n-gram cohesion and n-gram specificity guided by the latent concepts. The cost can be employed to re-rank the search results gener- ated from an information retrieval (IR) engine. The experimental results demonstrate that our framework achieves significant improvement in ranking documents against the state-of-the-art unsupervised comparative methods. KEYWORDS: Term Sequence, LSI, IR, Domain-specific Readability, Ranking, Dynamic program- ming. 1309 1 Introduction Readability assessment is an important issue in NLP (Tanaka-Ishii et al., 2010). Traditional general readability methods (Dubay, 2004) have been applied to several problem tasks such as matching books with grade levels (Collins-Thompson and Callan, 2005; Fry, 1969). However, the problem of readability has not been well explored in Information Retrieval (IR) (Kim et al., 2012). Recently researchers have started looking at the problem in IR and several motivations for incorporating readability in an IR system have already been clearly laid out in (Kim et al., 2012; Tan et al., 2012; Nakatani et al., 2009; Yan et al., 2006; Collins-Thompson et al., 2011; Jameel et al., 2011; Zhao and Kan, 2010; Kumaran et al., 2005). What has lacked is a thor- ough investigation into the problem of reading difficulty in domain-specific IR because tra- ditional unsupervised general readability formulae tend to underperform in domain-specific document retrieval (Yan et al., 2006; Jameel et al., 2011). Domain-specific IR is important be- cause many people are searching for information outside their domain of expertise (Bhavnani, 2002; Yan et al., 2006). The most affected group who regularly experience difficulties in retrieving documents based on readability are children (Collins-Thompson et al., 2011) and other users who are not do- main experts (Yan et al., 2006). Hence, most of them will look for domain-specific documents which they can easily comprehend (Bhavnani, 2002). Domain experts employ complex search strategies such as usage of jargon and complex phrases to successfully retrieve documents …

Content truncated to 500 words