Development And Evaluation of an Automatic Judgment Method Using Machine Learning for the Primary Screening of Historical Public Records

Abstract

Local governments face the problem of improving the efficiency of the selection process for historical public records. This study developed an automatic judgment method for the primary screening of historical public records by machine learning, using the administrative documents of Kumamoto Prefecture as a case study. Automatic determination was possible for 58% of all documents. Comparing the automatic judgment results with the actual manual judgment showed a recall rate of 60% and a precision rate of 32%. However, each rate has a trade-off relationship, and improving each index is difficult. Efforts to improve the accuracy of judgments are the following: （1） Selection of the most appropriate machine learning model exceeding the cosine similarity, （2） Introduction of a system to assign appropriate names to documents according to their contents, and （3） Reflection of the secondary screening results in the past years in the primary screening process.