In this paper we propose an efficient tech-nique for ranking triples of knowledgebase using information of full text. Wedevise supervised machine learning algo-rithms to compute the relevance scoresfor item-property pairs where an item canhave more than one value.Such a scoremeasures the degree to which an entitybelongs to a type, and this plays an im-portant role in ranking the search results.The problem is, in itself, new and not ex-plored so much in the literature, possiblybecause of the heterogeneous behaviors ofboth semantic knowledge base and full-text articles. The classifiers exploit statisti-cal features computed from the Wikipediaarticles and the semantic information ob-tained from the word embedding concepts.We develop models based on traditionalsupervised models like Suport Vector Ma-chine (SVM) and Random Forest (RF);and then using deep Convolution Neu-ral Network (CNN). We perform experi-ments as provided by WSDM cup 2017,which provides about 1k human judg-ments of person-profession pairs. Evalu-ation shows that machine learning basedapproaches produce encouraging perfor-mance with the highest accuracy of 71%.The contributions of the current work aretwo-fold,viz.we focus on a problem thathas not been explored much, and show theusage of powerful word-embedding fea-tures that produce promising results.