An Intelligent Matching Approach for Upstream and Downstream Textual Information in Manufacturing Supply Chains Based on Transformer

Abstract

In the era of intelligent manufacturing, supply chain collaboration has become increasingly dependent on the accurate and efficient exchange of information between upstream and downstream enterprises. However, the textual data exchanged in such processes—such as product descriptions, purchase orders, and specification sheets—are often unstructured and heterogeneous, posing significant challenges for semantic understanding and automated matching. Traditional keyword-based or statistical models fail to capture deep contextual relationships, resulting in inefficient coordination and high manual overhead. This paper proposes a novel Transformer-based intelligent matching framework designed specifically for the manufacturing industry's supply chain environment. The framework leverages the attention mechanisms and contextual embedding capabilities of Transformer models to achieve fine-grained semantic understanding between texts originating from different entities within the supply chain. The model accepts pairs of textual entries from upstream and downstream partners and computes a semantic similarity score that guides automated matching decisions.
To adapt the Transformer model to this domain, we introduce domain-specific pretraining and fine-tuning strategies using a curated dataset collected from real-world supply chain interactions. This dataset includes annotated text pairs representing successful and unsuccessful information matches, which are used to supervise the learning process. We also design a hybrid architecture that integrates domain knowledge features, such as part categories and business terminology, with the Transformer encoder to enhance model performance. Experiments are conducted on a newly constructed benchmark dataset containing over 10,000 annotated text pairs from various manufacturing sectors. Our proposed model is compared against traditional matching approaches, including TF-IDF, word embedding similarity, and deep Siamese networks, as well as baseline Transformer models such as BERT and RoBERTa. The results demonstrate that our model significantly outperforms existing methods in terms of accuracy, precision, recall, and F1-score, achieving a performance improvement of over 10% in key metrics. In addition to quantitative evaluation, we deploy the model in a prototype ERP-integrated recommendation system for matching supplier capabilities with procurement requirements. Real-world case studies validate the practical value of the system in reducing manual workload, improving match accuracy, and accelerating supply chain response time. This research contributes to the growing intersection of artificial intelligence and industrial supply chain management by offering a robust, scalable, and interpretable solution for semantic text matching. It paves the way for further integration of AI-driven decision support systems in manufacturing operations, particularly under the trend of digital transformation and smart factory initiatives. Future work will explore incorporating multimodal data such as diagrams and structured metadata, enhancing model interpretability via attention visualization, and extending the framework to multilingual and cross-cultural supply chain scenarios.