Automating Invoice Processing with Robotic Process Automation (RPA) and Large Language Model (LLM)
Open Access Depositedan automated process that combines robotic process automation (RPA) and a LLMs implemented locally. We divided this praxis into three parts, which are modeling, analysis, and implementation. Even though we used only around ten percent of the labeled data, the model still did well on key invoice fields, reaching about 100% accuracy, 68% recall, and an overall F1 score 0.81. In the value matching task, precision was 0.98, recall was 0.66, and F1 score was 0.79. An ablation study comparing models with and without knowledge distillation showed a significant improvement in performance, with the F1 score increasing by approximately threefold under the same training settings. Interestingly, the experiments also showed that optical character recognition (OCR), which is often considered necessary, is not always help. Since the models trained without OCR achieved slightly better results across several prompting methods. Overall, the results indicate that combining local LLMs with RPA and knowledge distillation can make invoice automation more efficient and privacy -preserving without heavy reliance on large annotated datasets.
For most organizations, invoice processing remains one of the most time-consuming parts of financial departments. Manual data review and entry is slow, error-prone, and difficult to scale as business volumes grow. While machine learning tools have been introduced to aid in the processing, they typically rely on a large number of labeled examples, which are expensive to create. Recent advances in large language models (LLMs) offer a way to reduce manual labeling without sacrificing too much accuracy. However, invoices contain sensitive financial information, which typically makes sending them to cloud-based commercial models unacceptable for privacy reasons. The goal of this research is to find a viable alternative
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.