With the growth in biomedical literature, the necessity of extracting useful information from the literature has increased. One approach to extracting biomedical knowledge involves using citation relations to discover entity relations. The assumption is that citation relations between any two articles connect knowledge entities across the articles, enabling the detection of implicit relationships among biomedical entities. The goal of this article is to examine the characteristics of biomedical entities connected via intermediate entities using citation relations aided by text mining. Based on the importance of symptoms as biomedical entities, we created triples connected via citation relations to identify drug–disease pairs with shared symptoms as intermediate entities. Drug–disease interactions built via citation relations were compared with co‐occurrence‐based interactions. Several types of analyses were adopted to examine the properties of the extracted entity pairs by comparing them with drug–disease interaction databases. We attempted to identify the characteristics of drug–disease pairs through citation relations in association with biomedical entities. The results showed that the citation relation‐based approach resulted in diverse types of biomedical entities and preserved topical consistency. In addition, drug–disease pairs identified only via citation relations are interesting for clinical trials when they are examined using BITOLA.

In today's era of information explosion, extracting entities and their relations in largescale, unstructured collections of text to better represent knowledge has emerged as a daunting challenge in biomedical text mining. To respond to the demand to automatically extract scientific knowledge with higher precision, the public knowledge discovery tool PKDE4J (Song et al., 2015) was proposed as a flexible text-mining tool. In this study, we propose an extended version of PKDE4J to represent scientific knowledge for literature-based knowledge discovery. Specifically, we assess the performance of PKDE4J in terms of three extraction tasks: entity, relation, and event detection. We also suggest applications of PKDE4J along three lines: (1) knowledge search, (2) knowledge linking, and (3) knowledge inference. We first describe the updated features of PKDE4J and report on tests of its performance. With additional options in the processes of named entity extraction, verb expansion, and event detection, we expect that the enhanced PKDE4J can be utilized for literature-based knowledge discovery.