Text categorization of nuclear system documents using semi-automatic approach

Abstract

The current practice of nuclear material export decision is neither time-efficient nor cost-effective. It is becoming more and more difficult for the current system, in which human experts manually evaluate the submitted documents and export details described therein for the purpose of export permission decision, to deal with the explosion of export review requests. Toward the improvement of the situation, this research pro-poses a new text categorization technique of nuclear system documents. The knowledge acquisition bottleneck arises from a small number of experts working on a large set of information and there have been a continuous stream of studies to address this issue through automatic or machine approaches. However, the automated machine approach had to compensate for its speed with the quality of work. In this research, we suggest and demonstrate a new evaluation system for nuclear export control with the categorization of the documents us-ing keyterms and preexistent information of the categories. In particular, for the extraction of keyterms, three alternative approaches were compared: (1) automatic keyword extraction approach using TF-IDF, (2) semi-automatic approach, in which the automatic keyword extraction results were reviewed and adjusted by student experts who majored in nuclear engineering, and (3) totally manual approach in which a very experienced senior field expert extracted keywords without any machine support. The study results show that the semi-automatic approach is the most efficient in categorizing keywords, even though it relies on the work of the student experts, suggesting that when it is utilized involving field experts, the text categorization results will be even further better. The combination of machine and human seems a promising solution that can success-fully reduce the knowledge acquisition bottle-neck with reduced time/cost and improved accuracy.