[5Yin-A] Poster 8

[5Yin-A-17] A Semantic Search System for National Tax Tribunal Decisions Based on Large Language Model Embeddings

〇Ryuya Yamagishi¹, Fumikatsu Anaguchi¹, Takeshi Morita¹ (1. AOYAMA GAKUIN UNIVERSITY)

Keywords:

Semantic Search,Large Language Model,Embeddings,National Tax Tribunal Decisions

In tax accounting practice, retrieving relevant past national tax tribunal decisions is essential for analyzing cases and organizing legal justifications. However, the Tax Account Information Network System (TAINS), which is widely used for searching national tax tribunal decisions, relies on keyword-based retrieval and often fails to retrieve appropriate documents due to variations in expressions and paraphrasing. This study proposes a semantic search system for national tax tribunal decisions based on Large Language Model embeddings. Each decision is decomposed into its title, summary, and body text according to its document structure, and each component is independently vectorized. During retrieval, search results obtained from multiple vector databases are aggregated at the document level to reflect the overall semantic relevance of each national tax tribunal decisions. In addition, a web-based search interface was implemented to support practical use, allowing users to review search results and provide relevance judgments. Comparative experiments with an existing keyword-based system and subjective evaluations by experienced TAINS users indicate that the proposed approach tends to rank more relevant national tax tribunal decisions higher, particularly for queries that produce a large number of search results and are difficult to narrow down using keywords alone.