Criteria-based Evidence Suggestion for Document Classification

Abstract

To make machine learning (ML) models and their outputs more reliable, eXplainable AI (XAI) has been actively discussed. In recent years, they successfully enchance the explanatory power of ML models and their outputs, and machine learning is contributing much to real-world operations that require more reliability. However, their explanability can be insufficient in real-world applications where a person reviews and classifies documents based on a explicit criteria. Consequently, a classification model still requires high human cost of checking its prediction outputs. In this study, we examine a method to construct a classification model that can be explained according to an explicit criteria described in text. Further, we present a method to provide evidence of its prediction outputs to make reviewers check efficient. We confirmed in the experiment that our approach is promising for reducing the checking cost.