Question Answering System for Legal Bar Examination Using Predicate Argument Structure

摘要

We developed a question answering system for legal bar exam, which can explain the way system solves based on underlying logical structures. We focus on the set of subject and object with their predicate, i.e. the predicate argument structure, in order to represent structures of legal documents. We implemented a couple of modules using different searching methods. Our system outputs results using these modules by learning each module's confidence value with SVM. We manually analyzed the difficulty level of the problems whether external knowledge is required or not. We created a structured synonym dictionary specialized to the legal domain, where predicates are categorized with their objects. This synonym dictionary could absorb superficial differences of predicates to solve the problems which do not require external knowledge. We confirmed that the system can solve more than 70% of simple problems. Our system achieved the second best score in Task 4 of the COLIEE 2018 shared task.