A Deep Learning Approach for Identifying Similarities in Contracted Object Description in Public Procurement

The digitization of public procurement has led to the automation of data analysis in public administration. However, using natural language in contract details still poses challenges in analysis and auditing. The large volume of unstructured data in bidding notices calls for tools that automate data analysis for citizens and auditing authorities. The description of the tender object is a critical problem in data analysis, as it lacks standardization and often contains unclear content. Detecting similarities between contract objects is crucial for analyzing public purchases and identifying fraudulent practices. However, manually reading all contracts to notice similarities is tedious and impractical. Therefore, automating the identification of similarities is essential. We propose a deep learning-based solution that uses the Euclidean distance between contextualized embeddings of contracted object descriptions to detect similarity. To improve the performance of the models, we used Named Entity Recognition (NER) to extract the central idea of the contract object, excluding confusing words and expressions. This approach improves the detection of textual similarity in public procurement objects. In general, these advances are designed to facilitate the timely analysis of public contracts.