CNN-LSTM-Based Bilingual Receipt Information Extraction Using Template-Based Data Generation

This paper discusses the development of a bilingual receipt information extraction system using a CNN-LSTM model with data augmentation techniques. The system targets the extraction of essential information such as company names, dates, and total amounts from receipts containing both Thai and English text. To address the limited availability of annotated data, synthetic receipt samples were generated from initial templates, creating a diverse training dataset. The model’s performance was evaluated on both the generated dataset and the SROIE 2019 dataset, achieving high accuracy across all tested information classes. While the CNN effectively extracts features, the LSTM processes these features for accurate information extraction. Future work aims to incorporate transformer-based models to enhance the system’s contextual understanding and generalization capabilities. This research highlights the effectiveness of combining CNNs and LSTMs in handling complex, multilingual datasets for practical applications in information extraction.