Concept Grounding to Multiple Knowledge Bases via Indirect Supervision

Concept Grounding to Multiple Knowledge Bases via Indirect Supervision Chen-Tse Tsai and Dan Roth University of Illinois at Urbana-Champaign 201 N. Goodwin, Urbana, Illinois, 61801 {ctsai12, danr}@illinois.edu Abstract We consider the problem of disambiguating concept mentions appearing in documents and grounding them in multiple knowledge bases, where each knowledge base addresses some aspects of the domain. This problem poses a few additional challenges beyond those ad- dressed in the popular Wikification problem. Key among them is that most knowledge bases do not contain the rich textual and structural information Wikipedia does; consequently, the main supervision signal used to train Wik- ification rankers does not exist anymore. In this work we develop an algorithmic approach that, by carefully examining the relations be- tween various related knowledge bases, gen- erates an indirect supervision signal it uses to train a ranking model that accurately chooses knowledge base entries for a given mention; moreover, it also induces prior knowledge that can be used to support a global coherent map- ping of all the concepts in a given document to the knowledge bases. Using the biomedical domain as our appli- cation, we show that our indirectly super- vised ranking model outperforms other unsu- pervised baselines and that the quality of this indirect supervision scheme is very close to a supervised model. We also show that consid- ering multiple knowledge bases together has an advantage over grounding concepts to each knowledge base individually. 1 Introduction Grounding entities and concepts appearing in text documents to a knowledge base (KB) has become a popular method for contextually disambiguating them and can be used also for focused knowledge acquisition. It has been shown a valuable component for several natural language processing and informa- tion extraction tasks across different domains. In the news domain, the task is often called Wikification or Entity Linking and has been studied extensively re- cently (Bunescu and Pasca, 2006; Cucerzan, 2007; Mihalcea and Csomai, 2007; Ratinov et al., 2011; Cheng and Roth, 2013). Wikipedia is widely used as the target KB due to its broad coverage and de- tailed information of concepts. While Wikipedia is an excellent general purpose encyclopedic resource, when the text is domain specific, it may not be the single ideal resource; the text could be better “cov- ered” by multiple ontological or encyclopedic re- sources. This is clearly the case for scientific text which is often covered by multiple ontologies, each ad- dressing some aspects of the domain. For exam- ple, in the biological domain there are multiple on- tologies: Entrez Gene (Lu and Wilbur, 2010) fo- cuses on genomes that have been completely se- quenced; Gene Ontology (Ashburner et al., 2000) more broadly describes gene product characteristics; and ChEBI, is a dictionary of molecular entities fo- cused on “small” chemical compounds. The ontolo- gies provide complementary information, but they overlap and, in these cases, make use of different vocabulary and provide different relevant informa- tion. In this paper, we consider the problem of ground- ing concepts appearing in documents to …

Content truncated to 500 words