HomeTax Comprehensive Income Tax Chatbot Demonstration

Overview

The National Tax Service operates a 'User Consultation Chatbot' for the convenience of HomeTax users, but its flexibility and accuracy in handling queries are limited due to the constraints of a standard rule-based chatbot. Therefore, to overcome the limitations of the existing model, this project aimed to enhance it into a chatbot capable of responding to given scenarios by training and validating the model using legal information.

Project description

This model was trained on 16,471 comprehensive income tax scenarios in CSV format held by the National Tax Service and on income tax law data. A RAG-based training dataset was constructed and fine-tuned by linking it with the HyperClova X model. Using this model, a chatbot demo page based on comprehensive income tax scenarios was built.

Key technological features

• End-to-End Integrated Implementation: Implementing an end-to-end system that integrates callbots with LLM-RAG systems to handle legal-based consultations
• Hybrid Search and Rerank Algorithm: Optimizing search performance through hybrid search experiments based on VectorDB and text similarity, and applying a rerank algorithm
• Graph DB-Based Reference Relationship Processing: Systematically processing reference relationships between legal data using a Graph DB and utilizing these relationships to enhance the Retriever
• Stream Processing: Improve real-time output response speed with voice bots by introducing Streamed Server Events (SSE)

Expected outcomes

When using the service for multi-turn conversations, hallucinations occur. Inaccuracies were in some responses to application questions, mitigated this by preprocessing some data and performing paragraph-level chunking. This service learns from highly complex, unstructured data like original legal statutes, enabling broader and more diverse responses than existing chatbots. Consequently, it improved customer satisfaction and demonstrated enhanced operational efficiency for staff. Furthermore, the LLM approach used in the service demonstrated a lower response failure rate compared to rule-based systems and exhibited superior performance in scenarios involving two or more mixed scenarios.

Launch year: 2023

Attribution

Case study courtesy of the National Information Society Agency, Republic of Korea, Apolitical's Content Partner. Sourced from Use Cases of Public AI Service, Vol.1,  https://eng.nia.or.kr/site/nia_eng/ex/bbs/View.do?cbIdx=31975&bcIdx=28997&parentSeq=29028