From natural language to accounting entries using a natural language processing method

Abstract

Bookkeeping is crucial in both accounting and auditing. However, a substantial quantity of accounting information initially recorded using unstructured natural language, which restricts the efficiency and accuracy of bookkeeping. In this study, we exploit proprietary transaction data from three firms to demonstrate the capacity of a word embedding approach based on a neural network model (i.e., Word2vec) for processing transaction‐related natural language and automating bookkeeping practice. Our study contributes to accounting practice and literature by demonstrating a practical application of Word2vec to the construction of an automated bookkeeping system.

Unleashing efficiency and insights: Exploring the potential applications and challenges of ChatGPT in accounting

“…Our study has significant implications for both scholars and accounting practitioners. For academic scholars, this study contributes to the emerging field of AI and NLP technologies in accounting (e.g., Bochkay et al., 2023; Chen et al., 2023; Cooper et al., 2022; Fisher et al., 2016) by shedding light on ChatGPT's capabilities and limitations and providing a foundation for further research in understanding GPT's impact on various accounting tasks. This research serves as a starting point for conducting empirical studies, theoretical analyses, and interdisciplinary research that combines NLP's iteration and accounting.…”

“…As one of the most popular techniques in the AI field, NLP technology has facilitated the extensive and efficient processing of unstructured textual materials by computers (e.g., Bochkay et al., 2023; Chen et al., 2023; Fisher et al., 2016). Generative Large Language Models (GLLMs), such as OpenAI's ChatGPT models, have emerged as powerful tools for textual analysis tasks in accounting research (De Kok, 2023).…”

Comparison of Prompt Engineering and Fine-Tuning Strategies in Large Language Models in the Classification of Clinical Notes

“…For example, Med-PaLM was a fine-tuned model from PaLM for biomedical research and performed considerably well but remained inferior to clinicians 9 . Compared to fine-tuned biomedical pretrained language models, LLMs like GPT-3.5 and GPT-4 showed promise in biomedical semantic similarity and reasoning tasks but are less effective in information extraction and classification 10 . Despite these efforts, few studies have systematically evaluated various prompting and fine-tuning strategies in published LLMs and compared with human performance.…”