РОЗРОБКА НАДІЙНОЇ ДІАЛОГОВОЇ СИСТЕМИ ДЛЯ ВИВЧЕННЯ ТА АНАЛІЗУ СТАНДАРТІВ ФІНАНСОВОГО ОБЛІКУ: КОНЦЕПЦІЇ ТА ВИКЛИКИ.

Abstract

Current paper provides challenges and concepts that emerge during development of ChatGPT-like question-answering system for analyze and study financial accounting standards (such as IFRS, US GAAP, UK GAAP, etc.) differencesand usage, including issues of compliance between different types of financial reporting in different standards, which calledFinancialStandardTableGPT (or FST-GPT). Issues of processing data in various types and formats, as well as various cases with their sources and initial quality were considered during system development. To solve these issues, it was proposed to use solutions based on consequent usage of Deep Learning models which are able to effectively process financial reporting information from alarge number of various formats because of this configuration. The paper also examines in detail the task of integrating data and commands formed on the basis of natural language and tabular data, including the issue of transforming natural language commands into a sequence of instructions over tabular data and final result forming, which is presented in the form of tabular data and an optional additional explanation, which is presented in natural language form. Inaddition, current paper examines the limitations of classic GPT architectures for working with tabular data and considers GPT-based solutions that allow to successfully solve this issue, their features and the specifics of forming the necessary dataset of tabular data, which is oriented towards the implementation in dialogue systems which using LLMs. As a result, proposed system is focused on processing of user queries in form of text combined with financial report tables in various formats. Architecture of the FST-GPT is based on set of Deep Learning and Machine Learning models used in succession and combined with fine-tuned domain-specific LLM