TY - CHAP

T1 - Enhancing Invoice Recognition with LLM Embeddings in GAT Networks

AU - Thiée, Lukas Walter

AU - Funk, Burkhardt

N1 - Publisher Copyright: Copyright © 2025 by Association for Information Systems (AIS). All rights reserved.

PY - 2025/8

Y1 - 2025/8

N2 - We propose a novel approach for invoice recognition by integrating Large Language Model Embeddings as semantic features into the nodes of a Graph Attention Neural Network. Both the language model and the graph structure provide rich contextual information for our model to enhance the classification of OCR tokens from invoice documents. The experimental results demonstrate improvements in the classification performance on our datasets by over 3%, highlighting the effectiveness of our multiple attention mechanism. The approach is transferable to all kinds of service systems that process visually rich documents.

AB - We propose a novel approach for invoice recognition by integrating Large Language Model Embeddings as semantic features into the nodes of a Graph Attention Neural Network. Both the language model and the graph structure provide rich contextual information for our model to enhance the classification of OCR tokens from invoice documents. The experimental results demonstrate improvements in the classification performance on our datasets by over 3%, highlighting the effectiveness of our multiple attention mechanism. The approach is transferable to all kinds of service systems that process visually rich documents.

KW - GAT

KW - Invoice Recognition

KW - LLM

KW - Business informatics

UR - http://www.scopus.com/inward/record.url?scp=105025356507&partnerID=8YFLogxK

M3 - Article in conference proceedings

AN - SCOPUS:105025356507

T3 - Americas Conference on Information Systems, AMCIS 2025

SP - 4483

EP - 4492

BT - Americas Conference on Information Systems, AMCIS 2025

PB - The Association for Information Systems (AIS)

T2 - 2025 Americas Conference on Information Systems, AMCIS 2025

Y2 - 14 August 2025 through 16 August 2025

ER -